A method and system for task path correction driven by reality feedback in AI agents

Abstract

This invention provides a task path correction method and system driven by real-world feedback for AI agents, relating to the field of artificial intelligence technology. The method includes: receiving a task objective and generating a task identifier, step identifier, and path version; recording completion criteria and executing task steps; receiving tool call results, user action feedback, expert review results, experimental results, approval rejections, incident samples, or external system feedback during or after execution; generating feedback identifiers and associating them with the task path; determining the execution status of task steps or task paths based on feedback results, completion criteria, and execution status; generating failed or rejected paths upon failure or rejection, forming suppression rules and future output changes; and writing the task path status, feedback results, failed paths, rejected paths, suppression rules, future output changes, or responsibility logs into a task path library for subsequent path selection, tool calls, suggested outputs, and execution strategy correction.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a task path correction method and system driven by real-world feedback for an AI agent. Background Technology

[0002] With the development of artificial intelligence technology, AI Agent systems are increasingly being applied to scenarios such as task planning, automated execution, tool invocation, document processing, code generation, data analysis, workflow collaboration, and enterprise office work. Existing AI Agents can understand task objectives based on user input and complete some complex tasks through task decomposition, path planning, and invoking external tools or systems. Compared to traditional intelligent question-answering systems, AI Agents possess a certain degree of autonomous planning and execution capabilities, and can coordinate across multiple steps to assist users in achieving complex goals.

[0003] In existing technologies, AI agent systems typically rely on large language models for task understanding and planning. They generate corresponding task execution flows based on user needs, breaking down complex objectives into multiple sub-tasks, and then completing the corresponding operations by calling tools such as search engines, databases, code execution environments, office software, knowledge base systems, or external interfaces. During task execution, some systems can record task context, execution status, and tool call results to support multi-round interactions and continuous task execution.

[0004] However, existing AI agent systems tend to equate task planning results or tool execution results with actual task completion results, lacking an effective mechanism to determine whether the actual goal has been achieved. When the actual execution result is inconsistent with the expected goal, existing systems struggle to adjust the original task path based on feedback information, leading to insufficient task execution reliability. Furthermore, existing AI agent systems typically only perform simple recording, failing to structure and save failure reasons, failure steps, and corresponding path versions. They also cannot further develop failed paths, rejected paths, and future execution inhibition rules. This results in the system potentially repeatedly generating or executing paths that have already been rejected by the actual results in similar tasks, reducing task execution efficiency and reliability. Summary of the Invention

[0005] To address the problem that existing AI Agent systems tend to equate task planning or tool execution results with actual task completion, lacking an effective mechanism to determine whether the actual goal has been achieved, and failing to adjust the original task path based on feedback when the actual execution result differs from the expected goal, leading to insufficient task execution reliability. Furthermore, existing AI Agent systems typically only record information in a simple manner, failing to structure and save failure reasons, failed steps, and corresponding path versions, nor can they further generate failed paths, rejected paths, and future execution suppression rules. This results in the system potentially repeatedly generating or executing paths that have already been rejected by the actual results in similar tasks, reducing task execution efficiency and reliability. Therefore, this invention provides an AI Agent task path correction method and system driven by real-world feedback.

[0006] The technical solutions provided by the embodiments of the present invention are as follows: First aspect: This invention provides an AI Agent reality feedback-driven task path correction method, the method being executed by at least one processor, comprising: Receive the task objective and generate a task identifier; A task path is generated based on the task objective. The task path includes step identifiers, stage status, and path version. Determine the completion criteria corresponding to the task objective or the task step, and associate the completion criteria with the task identifier or the corresponding step identifier; Execute task steps based on the task path and record the execution status of the task steps; Receive real-world feedback results during or after task execution; Generate a feedback identifier for the actual feedback result, and associate the feedback identifier with the task identifier, step identifier, or path version; Based on the actual feedback results, the completion criteria, and the execution status, the path status of the corresponding task step or task path is determined. When the path status is a failure status, a failure path is generated; when the path status is a rejected status, a rejected path is generated; when the path status is neither a failure status nor a rejected status, neither the failure path nor the rejected path is generated. In the event that the failed path or the rejected path is generated, a suppression rule is generated based on the failed path or the rejected path; Based on the actual feedback results, and combined with at least one of the generated failed paths, rejected paths, or suppression rules, future output changes are generated. Write the task path status, actual feedback results, failed paths, rejected paths, suppression rules, future output changes, path version changes, or responsibility logs to the task path library. During subsequent task path generation, path selection, tool invocation, suggested output, or execution strategy generation, the failed path, rejected path, suppression rule, or future output change are invoked to modify the subsequent task path or execution strategy.

[0007] The second aspect: This invention provides an AI Agent reality feedback-driven task path correction system, comprising: The task access and path construction module is used to receive the task target and generate the task identifier, and generate the task path based on the task target. The task path includes step identifier, stage status and path version. The execution and feedback receiving module is used to determine the completion criteria corresponding to the task objective or the task step, associate the completion criteria with the task identifier or the corresponding step identifier, execute the task steps based on the task path and record the execution status of the task steps, and receive real feedback results during or after task execution. The feedback association and status judgment module is used to generate a feedback identifier for the actual feedback result, associate the feedback identifier with a task identifier, step identifier or path version, and determine the path status of the corresponding task step or task path based on the actual feedback result, completion criteria and execution status; when the path status is a failure status, a failure path is generated; when the path status is a rejected status, a rejected path is generated; when the path status is neither a failure status nor a rejected status, no failure path or rejected path is generated. The suppression and output correction module is used to generate suppression rules based on the failed or rejected path when a failed path or rejected path is generated, and to generate future output changes based on the actual feedback results and at least one of the generated failed path, rejected path or suppression rules. The write-back and call-to-correct module is used to write the task path status, actual feedback results, failed paths, rejected paths, suppression rules, future output changes, path version changes or responsibility logs into the task path library, and call the failed paths, rejected paths, suppression rules or future output changes during subsequent task path generation, path selection, tool invocation, suggested output or execution strategy generation to correct the subsequent task path or execution strategy.

[0008] Third aspect: An embodiment of the present invention provides an AI Agent reality feedback-driven task path correction device, comprising: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it causes the at least one processor to perform the AIAgent reality feedback-driven task path correction method as described in the first aspect.

[0009] Fourth aspect: The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the AI ​​Agent reality feedback-driven task path correction method as described in the first aspect.

[0010] Fifth aspect: The present invention provides a computer program product, characterized in that: The computer program product includes a computer program that, when executed by a processor, implements the AI ​​Agent reality feedback-driven task path correction method as described in the first aspect.

[0011] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following: (1) By generating task identifiers for task objectives, setting step identifiers, stage statuses and path versions for task paths, determining the completion standards corresponding to task objectives or task steps, and associating the completion standards with task identifiers or corresponding step identifiers, recording the execution status and receiving real feedback results during task execution, and then judging whether the task steps or task paths are in a completed, failed, rejected, pending review, frozen or needing to be rolled back state based on the real feedback results, completion standards and execution status, so that the AI ​​Agent no longer uses only task planning results, step execution results or tool call results as the basis for task completion, but can combine real feedback results to judge whether the task objectives have been actually achieved, and identify and correct the task path status when the actual execution results are inconsistent with the expected objectives, thereby improving the accuracy of task execution result judgment and the reliability of task execution.

[0012] (2) By associating feedback identifiers with task identifiers, step identifiers, or path versions, failed paths and rejected paths are generated when a task step or task path is judged to have failed or been rejected. Furthermore, suppression rules and future output changes are generated. The task path status, actual feedback results, failed paths, rejected paths, suppression rules, future output changes, path version changes, or responsibility logs are written into the task path library, enabling the structured recording of failure reasons, failed steps, and corresponding path versions. During subsequent task path generation, path selection, tool invocation, suggested output, or execution strategy generation, the failed paths, rejected paths, suppression rules, or future output changes are invoked to correct subsequent task paths or execution strategies, thereby reducing the repeated generation or execution of failed or rejected paths in similar tasks and improving task execution efficiency and reliability. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a flowchart illustrating a task path correction method driven by reality feedback for an AI Agent, as provided in an embodiment of the present invention.

[0015] Figure 2 This invention provides a relationship diagram between task identifiers, step identifiers, and path versions.

[0016] Figure 3 This is a diagram showing the relationship between task execution results and real-world feedback, provided as an embodiment of the present invention.

[0017] Figure 4 This is a flowchart for determining the status of a feedback result-driven path, provided in an embodiment of the present invention.

[0018] Figure 5 This invention provides a graph showing the relationship between failed paths and rejected paths.

[0019] Figure 6 This is a diagram showing the relationship between a suppression rule and future output changes, provided as an embodiment of the present invention.

[0020] Figure 7 This invention provides a relationship diagram between a task path library, a feedback library, and a responsibility log write-back.

[0021] Figure 8 This is a flowchart of a subsequent task path correction and execution strategy update provided in an embodiment of the present invention.

[0022] Figure 9 This is an overall architecture diagram of an AI Agent reality feedback-driven task path correction system provided in an embodiment of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0024] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0025] Example 1 Reference manual attached Figure 1 The diagram illustrates a flowchart of a task path correction method driven by reality feedback for an AI Agent, provided by an embodiment of the present invention.

[0026] This invention provides a task path correction method driven by AI agent reality feedback. This method can be implemented by an AI agent reality feedback-driven task path correction device, which can be a terminal or a server. The processing flow of the AI ​​agent reality feedback-driven task path correction method may include the following steps: Receive the task objective and generate a task identifier; In some embodiments, the task objective includes at least one of user input task, system instruction task, business process task, project task, experimental task, document task, software development task, engineering task, approval task, or external system triggered task; the task path also includes at least one of target text, completion criteria, step sequence, path status, parent path version, rollback objective, or responsibility log.

[0027] Specifically, the AI ​​Agent receives user input, system instructions, business process requests, project goals, experimental goals, document tasks, software development tasks, engineering tasks, approval tasks, or tasks triggered by external systems, and generates a task identifier. This task identifier is used to identify a long-term task, a complex task, a project task, or a multi-step execution task.

[0028] Reference manual attached Figure 2 This diagram illustrates a relationship between task identifiers, step identifiers, and path versions provided in an embodiment of the present invention.

[0029] A task path is generated based on the task objective. The task path includes step identifiers, stage status, and path version. Specifically, the system can break down the task into multiple steps based on the task objective and generate a step identifier for each step. The system generates a path version for the initial path, which is used to record the version status of the task path and subsequent correction relationships.

[0030] Determine the completion criteria corresponding to the task objectives or task steps, and associate the completion criteria with the task identifier or the corresponding step identifier; In some embodiments, the completion criteria are used to distinguish between task plan generation, task step execution, successful tool invocation, and achievement of actual goals; The completion criteria include at least one of the following: output file generation, successful tool execution, experimental results meeting the standards, test passing, expert review passing, user confirmation, approval passing, external system confirmation, or actual results meeting the preset conditions. The execution status includes at least one of the following: not started, in execution, executed, completed, failed, rejected, pending review, frozen, or requiring rollback.

[0031] Execute task steps based on the task path and record the execution status of the task steps; Specifically, the AI ​​Agent executes task steps based on the task path and records the stage status. The stage status includes at least one of the following: not started, in progress, executed, completed, failed, rejected, pending review, frozen, or requiring rollback. During task execution, the system can call search tools, database tools, code tools, document tools, experimental equipment interfaces, engineering system interfaces, office system interfaces, business system interfaces, or external APIs.

[0032] Reference manual attached Figure 3 The diagram illustrates the correlation between task execution results and real-world feedback provided by an embodiment of the present invention.

[0033] Receive real-world feedback results during or after task execution; In some embodiments, the actual feedback results include at least one of tool call results, user action feedback, expert review results, experimental results, test results, approval rejections, accident samples, customer feedback, project review results, equipment return status, or external system feedback; associating the feedback identifier with the task identifier, step identifier, or path version also includes associating the feedback identifier with at least one of object identifier, claim identifier, variable identifier, tool call identifier, risk level, or responsibility log.

[0034] Optionally, when the tool call result indicates that the tool call was successful but the actual feedback result does not meet the completion criteria, the corresponding task step is marked as tool success but task incomplete; when the actual feedback result indicates tool failure, experiment failure, test failure, user denial, expert rejection, approval rejection, accident sample occurrence, external system anomaly, or actual result not meeting the criteria, the corresponding task step or task path is marked as failed, rejected, pending review, frozen, or needs to be rolled back.

[0035] Generate a feedback identifier for the actual feedback result, and associate the feedback identifier with the task identifier, step identifier, or path version; Specifically, the system generates a feedback identifier for the actual feedback result and associates the feedback identifier with at least one of the following: task identifier, step identifier, object identifier, claim identifier, variable identifier, tool call identifier, path version, or responsibility log.

[0036] Reference manual attached Figure 4 The diagram illustrates a feedback result-driven path status judgment flowchart provided by an embodiment of the present invention.

[0037] Reference manual attached Figure 5 This illustrates a graph showing the relationship between failed paths and rejected paths provided by an embodiment of the present invention.

[0038] Based on the actual feedback results, the completion criteria, and the execution status, the path status of the corresponding task step or task path is determined. When the path status is a failure status, a failure path is generated; when the path status is a rejected status, a rejected path is generated; when the path status is neither a failure status nor a rejected status, neither the failure path nor the rejected path is generated. It should be noted that when the corresponding task step or task path is not determined to be in a failed or rejected state, the system does not generate a failed or rejected path, but records and writes back the determined task path status; subsequent processing is executed according to the task path status, risk level, review requirements, or rollback target.

[0039] Specifically, based on feedback results, completion criteria, task objectives, and current stage status, the system determines whether the corresponding task step or task path is in a completed, failed, rejected, pending review, frozen, or rollback state. When a tool is successfully invoked but the completion criteria are not met, the system marks the step as "tool successful but task incomplete." When expert review is negative, user feedback is negative, experimental results are substandard, approval is rejected, or external system feedback is abnormal, the system marks the corresponding path as "failed," "rejected," "pending review," "frozen," or "rollback required." When the system determines that a task step or task path has failed, a failed path is generated. When the system determines that a task step or task path has been rejected by the user, expert, approver, external system, or actual results, a rejected path is generated. The failed or rejected path includes at least one of the following: relevant task identifier, step identifier, path version, feedback identifier, failure reason, rejection reason, applicable conditions, risk level, review requirements, and responsibility log.

[0040] Reference manual attached Figure 6 The diagram illustrates the relationship between a suppression rule and future output changes provided by an embodiment of the present invention.

[0041] In the event that the failed path or the rejected path is generated, a suppression rule is generated based on the failed path or the rejected path; It should be noted that the suppression rules are used to suppress repeated recommendations, repeated executions, repeated calls, or unconditional output of failed or rejected paths in future identical or similar tasks. The suppression rules may include at least one of the following: prohibiting calls, lowering priority, requiring explanation of failure history, requiring manual review, requiring confirmation of changes in conditions, or requiring the generation of alternative paths.

[0042] Based on the actual feedback results, and combined with at least one of the generated failed paths, rejected paths, or suppression rules, future output changes are generated. In some embodiments, the failure path is used to record task paths that have failed to execute, whose actual results are unsatisfactory, or whose completion criteria are not met. The failed path includes at least one of the following: task identifier, step identifier, path version, feedback identifier, failure reason, failure condition, scope of application, risk level, review requirement, or responsibility log; the rejected path is used to record task paths that are rejected by users, experts, approvers, customers, external systems, or actual feedback results, and the rejected path includes at least one of the following: task identifier, step identifier, path version, feedback identifier, rejection reason, rejection source, scope of application, risk level, review requirement, or responsibility log.

[0043] Write the task path status, actual feedback results, failed paths, rejected paths, suppression rules, future output changes, path version changes, or responsibility logs to the task path library. In some embodiments, the future output change is used to modify at least one of the following: subsequent task path generation, path priority, tool invocation, suggested output, risk warning, review requirements, or execution strategy. Reference manual attached Figure 7 This diagram illustrates the relationship between a task path library, a feedback library, and a responsibility log write-back, as provided in an embodiment of the present invention.

[0044] When a subsequent task meets the similar task condition with the failed or rejected path, the suppression rule or future output change is invoked. The similar task condition includes at least one of the following: similar task objectives, same objects, same steps, same tools, related path versions, same feedback types, same failure reasons, or same applicable conditions.

[0045] In some embodiments, the path version change includes generating a new path version while retaining the original path version, failed path, rejected path, or responsibility log, the new path version being used to record the corrected task path; when a task path is determined to be frozen or needs to be rolled back, a rollback target is generated, the rollback target pointing to at least one of the previous path version, previous step state, previous task state, or previous output state.

[0046] In some embodiments, the responsibility log is used to record at least one of the following processes: task target reception process, path version generation process, step execution process, tool invocation process, reality feedback reception process, path status judgment process, failed path generation process, rejected path generation process, suppression rule generation process, future output change generation process, write-back process, or invocation process. Optionally, the task path library includes at least one of the following: task path records, feedback records, failed path records, rejected path records, suppression rule records, future output change records, path version records, rollback target records, or responsibility log records.

[0047] Reference manual attached Figure 8 The diagram illustrates a flowchart of a subsequent task path correction and execution strategy update provided by an embodiment of the present invention.

[0048] During subsequent task path generation, path selection, tool invocation, suggested output, or execution strategy generation, the failed path, rejected path, suppression rule, or future output change are invoked to modify the subsequent task path or execution strategy.

[0049] Example 2 Scientific research experiment task feedback and correction scenario.

[0050] This embodiment illustrates how the present invention can be applied in scientific research experiment design, materials development, drug screening, engineering testing, algorithm experiments, or technology verification tasks. This embodiment focuses on how the AI ​​Agent writes failed experimental results back to the task path and generates failure paths, suppression rules, and future output changes accordingly.

[0051] In this embodiment, the user initiates a research task to the AI ​​Agent. For example, the user requests the system to assist in designing a set of material experimental schemes to verify whether a certain combination of parameters can improve material performance. After receiving the task objective, the system generates a task identifier and records the objective text and completion criteria. The completion criteria may include conditions such as completing the experimental scheme, completing the experimental execution, obtaining experimental data, performance indicators reaching a preset threshold, expert review approval, or user confirmation.

[0052] The system generates a task path based on the task objective and breaks the task down into multiple steps. The first step corresponds to collecting background information; the second step corresponds to determining candidate experimental parameters; the third step corresponds to generating an experimental plan; the fourth step corresponds to executing the experiment; the fifth step corresponds to analyzing the experimental results; and the sixth step corresponds to revising the experimental route based on the results. The system generates a path version for this task path.

[0053] During task execution, the AI ​​Agent generates the first experimental route based on historical data, literature sources, user input, and tool calculation results. This experimental route includes the experimental subjects, parameter combinations, sample conditions, testing methods, and expected metrics. The system then binds this route to task identifiers, step identifiers, object identifiers, and related claim identifiers.

[0054] After users perform the experiment according to the route, they report the results to the system. If the experimental data shows that the performance indicators do not reach the preset threshold, the system generates a feedback identifier and marks the feedback type as "experiment result feedback," and the actual result as "experiment failure" or "indicator not met." The system also records the experiment time, experiment conditions, sample number, test results, reason for failure, and user description.

[0055] The system associates feedback identifiers with task identifiers, fourth step identifiers, fifth step identifiers, path versions, experiment object identifiers, and related claim identifiers. Based on completion criteria, if the AI ​​Agent has generated an experiment plan and the user has completed the experiment, but the actual result did not meet the objective, then this task step cannot be considered complete and should be marked as failed or requiring correction.

[0056] The system then generates a failure path. This failure path includes a task identifier, step identifier, path version, feedback identifier, experimental parameters, experimental conditions, failure result, failure reason, applicable conditions, risk level, review requirements, and responsibility log. If the system determines that the failure only applies under the current parameter conditions, it limits the applicable conditions of the failure path to the current experimental conditions; if the system determines that the entire route is infeasible, it marks the relevant path status as failed and freezes it.

[0057] The system then generates suppression rules based on the failure path. These suppression rules are used to prevent the AI ​​Agent from repeatedly recommending the same parameter combination under the same or similar experimental conditions in the future. The suppression rules may include: no longer prioritizing the recommendation of this parameter combination under the same material system and the same process window; if the user requests to try again, the previous failure result and verification requirements must be displayed; and when generating alternative solutions, the system should prioritize avoiding parameter ranges that have failed.

[0058] The system further generates future output changes. These future output changes are used to modify the generation and suggested outputs of subsequent task paths. For example, when the AI ​​Agent generates an experimental plan, it should indicate that the route fails under the current conditions; it should recommend alternative experimental parameters; it should add verification steps; it should request supplementary experimental data; or it should reduce the original proposed output intensity.

[0059] Finally, the system writes the task path status, experimental feedback results, failed paths, suppression rules, future output changes, path version changes, and responsibility logs into the task path database, feedback database, failed path database, suppression rule database, and responsibility log database. Subsequently, when a user submits a similar experimental task, the AI ​​Agent invokes the aforementioned failed paths and suppression rules, and will no longer unconditionally recommend the same experimental route again.

[0060] Through this embodiment, the present invention enables the AI ​​Agent to absorb real experimental feedback in scientific research tasks, transform experimental failure results into a basis for task path correction, and avoid the system repeatedly recommending failed experimental routes.

[0061] Example 3 A scenario for revising patent drafting tasks driven by expert review and rejection.

[0062] This embodiment illustrates how the present invention can be applied to tasks such as patent drafting, technical disclosure, legal document generation, research report generation, or professional document generation. This embodiment focuses on how the AI ​​Agent writes expert review rejection opinions back into the task path and generates the rejected path and future output changes.

[0063] In this embodiment, the user requests an AI Agent to assist in completing patent disclosure materials for a specific technical solution. After receiving the task objective, the system generates a task identifier and records the completion criteria. These criteria may include the completed technical problem, technical solution, embodiments, draft claims, abstract, and accompanying drawings, and are subject to review by the user or an expert.

[0064] The system generates a task path. The first step corresponds to receiving technical documents; the second step corresponds to identifying core technical issues; the third step corresponds to breaking down the technical solution; the fourth step corresponds to generating draft claims; the fifth step corresponds to drafting embodiments; the sixth step corresponds to expert review; the seventh step corresponds to revising based on review comments; and the eighth step corresponds to forming the final draft. The system generates a path version for the initial path.

[0065] In the fourth step, the AI ​​Agent generates a set of draft claims containing a combination of technical features. Subsequently, expert review opinions indicate that the combination of technical features lacks novelty, highly overlaps with existing patents, or has an overly broad scope of protection, and therefore cannot serve as the core of the main claim. After receiving the expert review opinions, the system generates feedback identifiers, marking the feedback type as the expert review result and the actual result as expert rejection or failure.

[0066] The system then associates the feedback identifier with the task identifier, fourth step identifier, relevant object identifier, relevant claim identifier, path version, and responsibility log. Based on the completion criteria, the system determines that although a draft claim has been generated in this step, it has not passed expert review and therefore should not be marked as completed, but rather as rejected, pending amendment, or requiring rollback.

[0067] The system generates rejected paths. These rejected paths record the rejected combination of technical features, the corresponding claim paragraphs, the expert's reasons for rejection, relevant sources, risk level, review requirements, and a responsibility log. If the expert opinion indicates that the path cannot continue as the primary route, the system can mark it as a rejected primary route; if only narrowing the scope is required, the system can mark the path as needing modification.

[0068] The system generates suppression rules based on rejected paths. These suppression rules are used to prevent AIAgent from using the same combination of technical features as the core of the main claim in subsequent patent drafting tasks, unless the user or expert explicitly states that there are new technical differences, new implementation effects, or new supporting evidence. The system can also require subsequent outputs to indicate "This path was previously rejected by an expert."

[0069] The system generates future output variations. These future output variations include: avoiding the original rejected combination when generating subsequent claims; prioritizing the generation of alternative technical feature combinations; downgrading the original solution; treating it as a dependent implementation or alternative in the specification; adding technical effect limitations; or requiring further search and review.

[0070] The system can generate a second path version. This second path version retains the historical content, expert rejection records, and rejected paths from the first path version, while adding revised claim strategies and alternative routes. The system does not directly overwrite the old path, but rather preserves the inheritance relationships and rollback targets between path versions.

[0071] Finally, rejected paths, expert feedback, suppression rules, future output changes, and path version updates are written to the task path library, rejected path library, suppression rule library, and responsibility log library. When the AI ​​Agent continues to write code, it will call upon this record to avoid reverting to paths rejected by experts.

[0072] Through this embodiment, the present invention enables the AI ​​Agent to incorporate expert review results in professional document tasks, and to structure rejected routes into rejected paths and suppression rules, thereby improving long-term collaboration and professional correction capabilities.

[0073] Example 4 Software development agent tools execute feedback and correction scenarios.

[0074] This embodiment illustrates how the present invention can be applied in software development agents, code generation, automated testing, debugging tasks, continuous integration, or engineering automation scenarios. This embodiment focuses on how the system distinguishes between successful tool invocation and task completion, and writes test failure results back to the task path.

[0075] In this embodiment, the user requests the AI ​​Agent to develop a specific functional module. The system generates a task identifier and breaks down the task into steps such as requirements analysis, code generation, unit testing, integration testing, bug fixing, and user confirmation. The system generates a path version and records the completion criteria. Completion criteria may include code generation, successful testing, functional compliance with requirements, absence of critical errors, and user confirmation.

[0076] In the code generation step, the AI ​​Agent invokes a code generation tool to generate target code. Subsequently, the system invokes a testing tool to execute unit tests. The testing tool returns the execution results. If the testing tool invocation is successful, but the test results show multiple test cases failing, the system generates a tool feedback flag, marking the feedback type as a tool invocation result and the displayed result as a test failure.

[0077] The system associates feedback identifiers with task identifiers, test step identifiers, code object identifiers, path versions, and tool invocation identifiers. Based on completion criteria, a successful tool invocation only indicates that the test tool was successfully executed, not that the task is complete. If the test fails, the test step should be marked as failed or requires correction.

[0078] The system generates a failure path. This failure path records the failed test cases, error logs, code modules, failure reasons, triggering conditions, path versions, and responsibility logs. If the error originates from a code generation strategy, the system can record that strategy as part of the failure path.

[0079] The system generates suppression rules based on failure paths. For example, in subsequent code generation, the interface call method that caused the error will no longer be used, the same error pattern will not be generated repeatedly, or relevant test conditions must be checked before generating similar code. The system generates future output changes, including changing the code generation strategy, adding boundary condition checks, prioritizing the execution of relevant regression tests, and indicating the reasons for previous failures.

[0080] If the AI ​​Agent subsequently corrects its code and passes the test, the system generates a new path version, recording the transition from the failed path to the corrected path. The system retains the old path as a failed path and marks the new path as passed or pending user confirmation. If similar errors occur later, the system invokes the failed path and suppression rules to avoid generating the same type of error code repeatedly.

[0081] Through this embodiment, the present invention enables the software development agent to not only generate code and call tools, but also to correct the task path based on the test failure results, avoiding misjudging the successful tool call as the task completion.

[0082] Example 5 Workflow task correction scenario driven by enterprise approval rejection.

[0083] This embodiment illustrates how the present invention can be applied to enterprise processes, approval systems, smart office environments, contract processing, procurement applications, project initiation, or compliance review tasks. The focus is on how the AI ​​Agent writes approval rejections or process rejections back into the task path, creating a rejected path and future output changes.

[0084] In this embodiment, the AI ​​Agent assists the user in submitting a business process task, such as a procurement request, contract review, project initiation application, or compliance material submission. After receiving the task objective, the system generates a task identifier and a path version. The task steps include material collection, form completion, attachment organization, compliance check, submission for approval, approval feedback, and revised submission.

[0085] The AI ​​Agent completes the material preparation and submits it to the approval system. The approval system returns a result, showing that the approval has been rejected. Reasons for rejection include mismatched budget items, missing contract terms, incomplete attachments, inadequate approval authority, or insufficient compliance materials. The system generates an external feedback identifier, marking the feedback type as "approval rejected" and the actual result as "rejected" or "not approved."

[0086] The system associates feedback identifiers with task identifiers, submission approval step identifiers, related object identifiers, path versions, and responsibility logs. Based on completion criteria, if the AI ​​Agent has completed the submission action but the approval has not been granted, the task cannot be marked as completed; instead, it should be marked as rejected or pending correction.

[0087] The system generates a rejected path. This rejected path records the reason for the approval rejection, relevant fields, missing materials, approval nodes, applicable rules, and accountability logs. Based on this rejected path, the system generates suppression rules, such as requiring future similar processes to first check budget items, attachment completeness, and permission requirements; if identical materials are missing, the submission step should not be initiated directly; if the approval rules are not confirmed, the user should be prompted to supplement materials or undergo manual confirmation.

[0088] The system generates future output changes. When the AI ​​Agent generates similar approval materials subsequently, it will automatically remind the user to supplement previously missing materials, adjust form fields, add compliance check steps, or downgrade the submission action to "submit after confirmation." The system writes these results into the task path library, rejected path library, suppression rule library, and responsibility log library.

[0089] Through this embodiment, the present invention enables the AI ​​Agent to transform the enterprise's approval rejection results into a basis for subsequent process correction, avoiding repeated submission of unqualified materials and improving the efficiency of enterprise workflow execution.

[0090] Example 6 Task path correction scenario driven by on-site feedback in engineering projects.

[0091] This embodiment illustrates how the present invention can be applied to engineering project management, equipment installation, construction planning, industrial maintenance, quality inspection, or safety rectification tasks. This embodiment focuses on how the AI ​​Agent writes on-site feedback and incident samples back into the task path.

[0092] In this embodiment, the user requests the AI ​​Agent to assist in developing an equipment installation or engineering rectification plan. The system generates a task identifier and records the target text, completion criteria, risk level, and path version. The task path includes data verification, solution generation, tool calculation, on-site execution, quality inspection, on-site feedback, and rectification correction.

[0093] The AI ​​Agent generates installation paths or rectification plans based on the data. After execution, on-site personnel provide feedback: a certain construction step cannot be implemented, a certain equipment parameter does not match the on-site conditions, or the rectification still fails the quality inspection. The system generates a user action feedback indicator, recording the actual result as on-site execution failure or quality inspection failure.

[0094] If the feedback involves security risks, the system will raise the risk level and mark the relevant path status as failed, frozen, or pending review. The system generates failed paths, recording the on-site conditions, failed steps, execution results, failure reasons, risk level, and review requirements. If an incident sample appears, the system can associate the incident sample with the task identifier, step identifier, and path version, and trigger a rollback target.

[0095] The system generates suppression rules based on failure paths. For example, in the future, under the same site conditions, the same equipment type, or the same risk level, this rectification step will no longer be directly recommended; site parameters must be verified first; a safety review node must be added; or it must be implemented only after confirmation by a human engineer.

[0096] The system generates future output changes, enabling the AI ​​Agent to proactively prompt the reasons for previous on-site failures, applicable conditions, and safety limitations when outputting subsequent engineering solutions. The system writes failure paths, incident samples, suppression rules, freeze states, rollback targets, and responsibility logs into the task path library and risk library.

[0097] Through this embodiment, the present invention enables the AI ​​Agent to correct the task path in engineering projects based on on-site feedback, avoiding the continued recommendation of paths that are unsuitable for on-site conditions or pose safety risks.

[0098] Example 7 Feedback from long-term personal project assistants to correct scenarios.

[0099] This embodiment illustrates how the present invention can be applied in personal AI assistants, long-term writing assistants, learning assistants, research assistants, project management assistants, or long-term collaboration systems. This embodiment focuses on how the system writes user denial, changes in user preferences, and changes in project stages back into the task path.

[0100] In this embodiment, the user uses the AI ​​Agent to advance a writing, learning, or research project over a long period. The system generates a task identifier for the project and breaks down the task into steps such as data organization, outline generation, chapter writing, proofreading, review, archiving, and publishing. The system records the path version and completion criteria.

[0101] At one stage, the AI ​​Agent repeatedly outputs content according to the old format. User feedback includes: "Don't use this path anymore," "I've rejected this direction," "Don't write it like this again," and "This version is outdated." The system generates user feedback identifiers, marking the feedback type as user rejection or user preference update, and recording the actual result as rejected or discontinued.

[0102] The system associates this feedback identifier with a task identifier, step identifier, path version, related object identifier, or variable identifier. Based on the feedback content, the system determines whether the original path should be marked as a rejected path, a frozen path, or a path that needs to be rolled back. If the user explicitly requests that the path will no longer be used, the system generates a rejected path and sets suppression rules.

[0103] The system generates future output changes, causing subsequent outputs to be executed according to the new path, format, or stage. The system writes user feedback, rejected paths, suppression rules, and future output changes into the task path library and responsibility log library.

[0104] Through this embodiment, the present invention enables a personal AI Agent to absorb user denials and changes in project status during long-term collaboration, avoiding repeated contamination of subsequent task outputs by old paths.

[0105] Example 8 System deployment and data storage methods.

[0106] This embodiment illustrates one deployment method of the present invention in a computer system.

[0107] In one implementation, the system of the present invention includes a task input layer, a task path layer, an execution record layer, a feedback receiving layer, a path correction layer, an output strategy layer, and a data storage layer.

[0108] The task input layer receives user tasks, system instructions, business process requests, or tasks triggered by external systems. The task path layer generates task identifiers, step identifiers, path versions, and completion criteria. The execution record layer records the AIAgent's task execution steps, tool calls, knowledge base invocations, output generation, and entry into external systems. The feedback receiving layer receives tool call results, user action feedback, expert review results, experimental results, approval rejections, incident samples, or external system feedback. The path correction layer determines path status and generates failed paths, rejected paths, suppression rules, future output changes, and rollback targets. The output strategy layer invokes suppression rules and future output changes in subsequent tasks, altering path generation, tool calls, suggested outputs, and risk warnings. The data storage layer stores the task path library, feedback library, failed path library, rejected path library, suppression rule library, risk library, and responsibility log library.

[0109] In a data structure, the task table should at least include a task identifier, target text, completion criteria, task type, risk level, and responsibility log. The step table should at least include a step identifier, task identifier, step name, stage status, path version, and execution record. The path version table should at least include the path version, parent path version, reason for correction, rollback target, and update time. The feedback table should at least include a feedback identifier, feedback type, actual result, feedback source, feedback time, related tasks, related steps, and related path versions. The failed path table should at least include a failed path identifier, task identifier, step identifier, path version, reason for failure, applicable conditions, and review requirements. The rejected path table should at least include a rejected path identifier, reason for rejection, source of rejection, scope of application, and responsibility log. The suppression rule table should at least include a suppression rule identifier, applicable conditions, suppression object, suppression action, triggering condition, and failure condition. The future output change table should at least include a change identifier, affected task type, output restrictions, path priority change, risk warning change, and review requirement change.

[0110] The data described above can be implemented using relational databases, graph databases, document databases, vector databases, object storage, log libraries, or a hybrid storage approach. Fields can be linked using primary keys, foreign keys, graph relationships, indexes, timestamps, task identifiers, step identifiers, path versions, or responsibility logs.

[0111] Through this embodiment, the present invention can be deployed as an independent AI Agent task feedback closed-loop system, middleware, enterprise private platform, cloud service, API service, plug-in system or local software module, and can be integrated with existing AI Agents, workflow systems, knowledge bases, business systems and toolchains.

[0112] Reference manual attached Figure 9 The diagram shows the overall architecture of an AI Agent reality feedback-driven task path correction system provided by the present invention.

[0113] This invention also provides an AI Agent reality feedback-driven task path correction system, comprising: The task access and path construction module is used to receive the task target and generate the task identifier, and generate the task path based on the task target. The task path includes step identifier, stage status and path version. The execution and feedback receiving module is used to determine the completion criteria corresponding to the task objective or the task step, associate the completion criteria with the task identifier or the corresponding step identifier, execute the task steps based on the task path and record the execution status of the task steps, and receive real feedback results during or after task execution. The feedback association and status judgment module is used to generate a feedback identifier for the actual feedback result, associate the feedback identifier with a task identifier, step identifier or path version, and determine the path status of the corresponding task step or task path based on the actual feedback result, completion criteria and execution status; when the path status is a failure status, a failure path is generated; when the path status is a rejected status, a rejected path is generated; when the path status is neither a failure status nor a rejected status, no failure path or rejected path is generated. The suppression and output correction module is used to generate suppression rules based on the failed or rejected path when a failed path or rejected path is generated, and to generate future output changes based on the actual feedback results and at least one of the generated failed path, rejected path or suppression rules. The write-back and call-to-correct module is used to write the task path status, actual feedback results, failed paths, rejected paths, suppression rules, future output changes, path version changes or responsibility logs into the task path library, and call the failed paths, rejected paths, suppression rules or future output changes during subsequent task path generation, path selection, tool invocation, suggested output or execution strategy generation to correct the subsequent task path or execution strategy.

[0114] The AI ​​Agent reality feedback driven task path correction system provided by this invention can execute the AI ​​Agent reality feedback driven task path correction method described above and achieve the same or similar technical effects. To avoid duplication, this invention will not elaborate further.

[0115] The above are merely specific embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. The scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A task path correction method driven by real-world feedback for an AI agent, characterized in that, The method is executed by at least one processor and includes: Receive the task objective and generate a task identifier; A task path is generated based on the task objective. The task path includes step identifiers, stage status, and path version. Determine the completion criteria corresponding to the task objective or the task step, and associate the completion criteria with the task identifier or the corresponding step identifier; Execute task steps based on the task path and record the execution status of the task steps; Receive real-world feedback results during or after task execution; Generate a feedback identifier for the actual feedback result, and associate the feedback identifier with the task identifier, step identifier, or path version; Based on the actual feedback results, the completion criteria, and the execution status, the path status of the corresponding task step or task path is determined. When the path status is a failure status, a failure path is generated; when the path status is a rejected status, a rejected path is generated; when the path status is neither a failure status nor a rejected status, neither the failure path nor the rejected path is generated. In the event that the failed path or the rejected path is generated, a suppression rule is generated based on the failed path or the rejected path; Based on the actual feedback results, and combined with at least one of the generated failed paths, rejected paths, or suppression rules, future output changes are generated. Write the task path status, actual feedback results, failed paths, rejected paths, suppression rules, future output changes, path version changes, or responsibility logs to the task path library. During subsequent task path generation, path selection, tool invocation, suggested output, or execution strategy generation, the failed path, rejected path, suppression rule, or future output change are invoked to modify the subsequent task path or execution strategy.

2. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The task objectives include at least one of the following: user input tasks, system instruction tasks, business process tasks, project tasks, experimental tasks, document tasks, software development tasks, engineering tasks, approval tasks, or external system-triggered tasks; the task path also includes at least one of the following: target text, completion criteria, step sequence, path status, parent path version, rollback target, or responsibility log.

3. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The completion criteria are used to distinguish between task plan generation, task step execution, successful tool invocation, and achievement of actual goals; The completion criteria include at least one of the following: output file generation, successful tool execution, experimental results meeting the standards, test passing, expert review passing, user confirmation, approval passing, external system confirmation, or actual results meeting the preset conditions. The execution status includes at least one of the following: not started, in execution, executed, completed, failed, rejected, pending review, frozen, or requiring rollback.

4. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The actual feedback results include at least one of the following: tool call results, user action feedback, expert review results, experimental results, test results, approval rejections, accident samples, customer feedback, project review results, equipment return status, or external system feedback; associating the feedback identifier with the task identifier, step identifier, or path version also includes associating the feedback identifier with at least one of the following: object identifier, claim identifier, variable identifier, tool call identifier, risk level, or responsibility log.

5. The AI ​​Agent reality feedback-driven task path correction method according to claim 4, characterized in that, When the tool call result indicates that the tool call was successful but the actual feedback result does not meet the completion criteria, the corresponding task step will be marked as tool success but task incomplete; when the actual feedback result indicates tool failure, experiment failure, test failure, user denial, expert rejection, approval rejection, accident sample occurrence, external system anomaly, or actual result not meeting the criteria, the corresponding task step or task path will be marked as failed, rejected, pending review, frozen, or need to be rolled back.

6. The AI ​​Agent reality feedback-driven task path correction method according to claim 1, characterized in that, The failure path is used to record the task path that has failed to execute, whose actual result does not meet the standard, or whose completion criteria are not met. The failed path includes at least one of the following: task identifier, step identifier, path version, feedback identifier, failure reason, failure condition, scope of application, risk level, review requirement, or responsibility log; the rejected path is used to record task paths that are rejected by users, experts, approvers, customers, external systems, or actual feedback results, and the rejected path includes at least one of the following: task identifier, step identifier, path version, feedback identifier, rejection reason, rejection source, scope of application, risk level, review requirement, or responsibility log.

7. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The suppression rules are used to suppress repeated recommendations, repeated executions, repeated calls, or unconditional output of the failed or rejected paths in subsequent identical or similar tasks; the suppression rules include at least one of the following: prohibiting calls, reducing priority, requiring failure history to be displayed, requiring rejection history to be displayed, requiring manual review, requiring confirmation of changes in conditions, requiring generation of alternative paths, or requiring risk warnings.

8. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The future output changes are used to modify at least one of the following: subsequent task path generation, path priority, tool invocation, suggested output, risk warning, review requirements, or execution strategy. When a subsequent task meets the similar task condition with the failed or rejected path, the suppression rule or future output change is invoked. The similar task condition includes at least one of the following: similar task objectives, same objects, same steps, same tools, related path versions, same feedback types, same failure reasons, or same applicable conditions.

9. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The path version change includes generating a new path version while retaining the original path version, failed path, rejected path, or responsibility log. The new path version is used to record the corrected task path. When a task path is determined to be frozen or needs to be rolled back, a rollback target is generated. The rollback target points to at least one of the previous path version, previous step status, previous task status, or previous output status.

10. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The responsibility log is used to record at least one of the following processes: task target reception process, path version generation process, step execution process, tool invocation process, reality feedback reception process, path status judgment process, failed path generation process, rejected path generation process, suppression rule generation process, future output change generation process, write-back process, or invocation process; The task path library includes at least one of the following: task path records, feedback records, failed path records, rejected path records, suppression rule records, future output change records, path version records, rollback target records, or responsibility log records.

11. The AI ​​Agent Reality Feedback-Driven Task Path Correction Method according to claim 1, characterized in that, The method further includes: Generate a task path correction report, which includes at least one of the following: task identifier, step identifier, path version, actual feedback result, path status, failed path, rejected path, suppression rule, future output change, rollback target, or responsibility log.

12. A task path correction system driven by real-world feedback for an AI agent, characterized in that, include: The task access and path construction module is used to receive the task target and generate the task identifier, and generate the task path based on the task target. The task path includes step identifier, stage status and path version. The execution and feedback receiving module is used to determine the completion criteria corresponding to the task objective or the task step, associate the completion criteria with the task identifier or the corresponding step identifier, execute the task steps based on the task path and record the execution status of the task steps, and receive real feedback results during or after task execution. The feedback association and status judgment module is used to generate a feedback identifier for the actual feedback result, associate the feedback identifier with a task identifier, step identifier or path version, and determine the path status of the corresponding task step or task path based on the actual feedback result, completion criteria and execution status; when the path status is a failure status, a failure path is generated; when the path status is a rejected status, a rejected path is generated; when the path status is neither a failure status nor a rejected status, no failure path or rejected path is generated. The suppression and output correction module is used to generate suppression rules based on the failed or rejected path when a failed path or rejected path is generated, and to generate future output changes based on the actual feedback results and at least one of the generated failed path, rejected path or suppression rules. The write-back and call-to-correct module is used to write the task path status, actual feedback results, failed paths, rejected paths, suppression rules, future output changes, path version changes or responsibility logs into the task path library, and call the failed paths, rejected paths, suppression rules or future output changes during subsequent task path generation, path selection, tool invocation, suggested output or execution strategy generation to correct the subsequent task path or execution strategy.

13. A task path correction device driven by real-world feedback for an AI agent, characterized in that, include: At least one processor; and a memory communicatively connected to the at least one processor; The memory stores a computer program that can be executed by the at least one processor. When the computer program is executed by the at least one processor, it causes the at least one processor to perform the AI ​​Agent reality feedback-driven task path correction method as described in any one of claims 1 to 11.

14. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program that, when executed by a processor, implements the AI ​​Agent reality feedback-driven task path correction method as described in any one of claims 1 to 11.

15. A computer program product, characterized in that: The computer program product includes a computer program that, when executed by a processor, implements the AI ​​Agent reality feedback-driven task path correction method as described in any one of claims 1 to 11.

Images