Method and system for multi-agent collaboration and dynamic DAG arrangement of large model for FICC intelligent research and computer readable storage medium

Abstract

The application discloses a large model multi-agent cooperation and dynamic DAG arrangement method and system for FICC intelligent investment and research and a computer readable storage medium, and belongs to the technical field of artificial intelligence. The method comprises the following steps: analyzing logical dependency constraints among multiple agents, calculating the initial in-degree of a node, and constructing an initial DAG; scheduling distributed multi-agent nodes, performing financial investment and research reasoning and calculation tasks in parallel according to the in-degree priority of the DAG; monitoring the output data of each agent node in real time, calculating the information entropy score, and triggering a dynamic fuse mechanism; suspending the original downstream node of the current agent node, instantiating a remedial agent node, and reconstructing the memory allocation of the calculation graph by the remedial agent node inheriting the context memory pointer of the suspended node; and outputting a structured decision conclusion with a causal root hash value. The application solves the technical problems of black box illusion of large models in the financial field and static workflow that cannot be dynamically and adaptively expanded.

Description

Technical Field

[0001] This invention belongs to the technical field of interdisciplinary business of artificial intelligence and financial technology, specifically involving a method, system and computer-readable storage medium for large-scale multi-agent collaboration and dynamic DAG orchestration for intelligent investment research of FICC. Background Technology

[0002] In intelligent investment research for FICC (Fixed Income, Foreign Exchange, and Commodities), Large Language Model (LLM) systems suffer from the following serious problems: 1. Black box problem and data illusion; models lacking strict constraints often produce logical gaps.

[0003] 2. Existing workflows mostly use static DAGs, which cannot adaptively expand the topology based on abnormal intermediate results generated by inference.

[0004] 3. Not only does it waste high computing power, but the static black-box architecture also makes it difficult to meet the stringent compliance audit and traceability requirements of financial institutions. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and computer-readable storage medium for large-scale multi-agent collaboration and dynamic DAG orchestration in FICC intelligent investment research, so as to solve the black box illusion problem in the prior art.

[0006] To achieve the above objectives, this invention provides a method for large-scale multi-agent collaboration and dynamic DAG orchestration in FICC intelligent investment research, comprising the following steps: Step S101: parse the logical dependency constraints between multiple agents, calculate the initial in-degree of the nodes, and construct the initial dynamic directed acyclic graph topology. Step S102: Schedule distributed multi-agent nodes to execute financial investment research reasoning and calculation tasks in parallel according to the in-degree priority of the dynamic directed acyclic graph topology; Step S103: Monitor the output data of each intelligent agent node in real time, calculate the information entropy score of the output probability distribution of the intelligent agent node, and calculate the confidence level by matching the preset logical verification rules. If the information entropy score is higher than the set entropy threshold or the confidence level is lower than the preset confidence threshold, trigger the dynamic circuit breaker mechanism. Step S104: The dynamic circuit breaker mechanism suspends the original downstream node of the current agent node, instantiates a remedy agent node, and the remedy agent node inherits the context memory pointer of the suspended node to reconstruct the memory allocation of the computation graph. Step S105: Extract the input hash of each agent node, run the Prompt, output the result and microsecond-level timestamp, construct the Merkle tree structure and calculate the global causal root hash value, and output the structured decision conclusion with the causal root hash value.

[0007] Preferably, before step S101, the method further includes: receiving a multi-source heterogeneous FICC data stream containing macroeconomic indicators, bond spreads, and tick-level transaction data from the market data middleware and macro data interface, and structuring and cleaning it into a columnar structure with microsecond timestamps according to a unified mode, as the underlying context input data of the dynamic directed acyclic graph topology.

[0008] Preferably, step S101 includes maintaining a node table and an adjacency table in memory; each record in the node table includes a node number, Agent type, in-degree counter, and state machine field, and the state machine values ​​include five types: waiting, running, suspended, reviewed, and circuit breaker; the adjacency table records a list of direct downstream node numbers for each node; the ready queue is implemented as a lock-free first-in-first-out queue.

[0009] Preferably, the node table and adjacency table are maintained in memory, which includes the establishment of directed edges based on the strong input-output dependencies between agents: after reading the dependency configuration, the adjacency table is generated, the in-degree of each node is counted to obtain the in-degree table, and N1 with an in-degree of 0 is pushed into the ready queue.

[0010] Preferably, the scheduling process in step S102 includes: Push the agent node with an in-degree of 0 into the ready queue, allocate GPU / CPU computing resources, and trigger execution; After the current agent node finishes execution, the in-degree of its direct downstream node is decremented by 1; if the in-degree of the downstream node decreases to 0, it is pushed back into the ready queue for scheduling.

[0011] Preferably, the specific process of inheriting the context memory pointer of the suspended node in step S104 includes: Read the memory starting address and context state handle of the suspended node, and perform concurrency protection through read-write locks and reference counting mechanisms; The memory starting address is directly mapped to the input buffer of the newly instantiated remedial agent node, achieving zero-copy memory sharing.

[0012] Preferably, the remedial execution process is as follows: the remedial verification agent node N3' reads the upstream intermediate result in the shared memory, replaces the verification prompt and lowers the sampling temperature parameter before regenerating, and executes step S103 again; after the verification is passed, the output of N3' is reconnected to the original downstream N4 to restore the scheduling of the suspended branch.

[0013] Preferably, after triggering the dynamic circuit breaker mechanism, an abnormal retry closed-loop strategy is also included: if the output information entropy of the remedial agent node is still higher than the threshold after a limited number of retries, a system-level degradation strategy is triggered to terminate the graph computation of the computation branch in advance and mark the alarm log.

[0014] This invention also provides a large-scale multi-agent collaboration and dynamic DAG orchestration system for intelligent FICC investment research, implemented based on the above method, comprising: The data input module is used to receive FICC multi-source data streams; A dynamic DAG orchestration engine is used to resolve agent dependencies and maintain dynamic topology; The distributed intelligent agent execution cluster includes macro-analysis intelligent agents, factor calculation intelligent agents, and risk control and compliance intelligent agents running in parallel; The dynamic rerouting and logical circuit breaker module is used to calculate the information entropy score and schedule the remedial agent node and memory pointer inheritance when the circuit breaker is triggered; Merkle tree white-box causal chain persistence module, used to generate and store Merkle root hash values; The decision output terminal is used to display the conclusions with causal root hash values.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the method as described in any of the preceding claims.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: It improves parallel efficiency through an in-degree decreasing scheduling mechanism; it uses a large model information entropy circuit breaker mechanism to specifically intercept logical illusions; it implements zero-copy inheritance of context memory pointers to achieve zero-copy interpolation and anomaly closure for the remedial agent; and finally, it constructs the execution snapshot as a Merkle tree for evidence storage, ensuring 100% tamper-proof computation and eliminating compliance audit blind spots in intelligent investment research. This invention combines four technical means to overcome the black-box illusion problem of traditional large models in financial scenarios, and solves the technical bottleneck of static workflows not being able to dynamically and adaptively expand through zero-copy inheritance of memory pointers, achieving highly reliable intelligent investment research decision-making with white-box traceability. Attached Figure Description

[0017] Figure 1 This is an architecture diagram of one embodiment of the system of the present invention; Figure 2 This is a flowchart illustrating the steps of one embodiment of the method of the present invention; Figure 3 This is a schematic diagram of the initial DAG topology and in-degree scheduling in an embodiment of the method of the present invention; Figure 4 This is a schematic diagram of dynamic reconstruction after circuit breaking in an embodiment of the method of the present invention; Figure 5 This is a schematic diagram of zero-copy inheritance of context memory pointers in an embodiment of the method of the present invention; Figure 6 This is a schematic diagram of the Merkle tree causal chain evidence storage structure in an embodiment of the method of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort, and the details in this specification, can also be modified and changed based on different viewpoints and applications without departing from the spirit of the present invention. All modifications and embodiments fall within the protection scope of the present invention.

[0019] Related technical concepts involved in this invention: FICC: An abbreviation for Fixed Income, Currency, and Commodity. In the international financial market, it specifically refers to the business area that covers these three asset classes and is often referred to as "general fixed income".

[0020] Agent: refers to an independently deployed inference process that encapsulates a large language model instance and a role Prompt.

[0021] Prompt: Input instructions or questions received by the Agent.

[0022] DAG (Directed Acyclic Graph): A graph structure consisting of nodes and directed edges, where the direction of all edges is clear and there are no cyclic paths.

[0023] In-degree and ready queue: In-degree refers to the number of direct upstream dependencies of a DAG node that have not yet been completed; the ready queue stores nodes with an in-degree of 0 that can be executed immediately.

[0024] Information entropy: The probability distribution p(x) of candidate tokens output by the softmax layer of the model is calculated as H(X)=-∑p(x)log2p(x); the higher the entropy value, the more the model hesitates among multiple candidate conclusions, and the greater the risk of illusion.

[0025] Context memory pointer: refers to the handle to the starting address of the intermediate result memory block (including KV-Cache and structured intermediate conclusions) produced by the suspended node.

[0026] Merkle Trees and WORM: A Merkle Tree is a binary tree that hashes by concatenating leaf hashes pairwise upwards; WORM (Write-Once-Read-Many) is an immutable storage medium that allows for writing once and reading many times.

[0027] like Figure 1 As shown, this invention provides an embodiment of a large-scale multi-agent collaboration and dynamic DAG orchestration system for intelligent FICC investment research. The system includes: Data access module 1 receives FICC multi-source heterogeneous data streams. Specifically, module 1 connects to the market data middleware and the macro data interface, and cleans the FICC multi-source heterogeneous data streams, which contain macroeconomic indicators, bond spreads, and tick-level transaction data, into a columnar structure with microsecond timestamps according to a unified pattern. This structure is then written into the context data base as the underlying context of the DAG for each agent to read.

[0028] The Dynamic DAG Topology Orchestration Engine 2 parses Agent dependencies and maintains dynamic topologies. Regarding data structure maintenance, the Dynamic DAG Topology Orchestration Engine 2 constructs and maintains a node table, adjacency table, and ready queue in memory. Each record in the node table contains a node number, Agent type, in-degree counter, and state machine fields. The state machine values ​​include five states: waiting, running, suspended, reviewed, and circuit breaker. The adjacency table records a list of direct downstream node numbers for each node. The ready queue is implemented as a lock-free, first-in-first-out queue. In terms of topology generation and scheduling, the dynamic DAG topology orchestration engine 2 establishes directed edges based on the strong input-output dependencies between agents (for example, if factor calculation must wait for macroscopic analysis results, then a directed edge N1→N3 is established). Specifically, after reading the dependency configuration, the dynamic DAG topology orchestration engine 2 generates an initial adjacency list {N1:[N2,N3], N2:[N3], N3:[N4,N5], N4:[N5]}, and iterates through and counts the number of direct predecessors of each node to obtain an initial in-degree list {N1:0,N2:1, N3:2, N4:1, N5:2}. Finally, the first node N1 with an in-degree of 0 is pushed into the ready queue, triggering parallel scheduling.

[0029] The distributed agent execution cluster 3, in this embodiment, comprises three processes running independently on the operating system: macro analysis agent 3a, factor calculation agent 3b, and risk control and compliance agent 3c. These processes exchange scheduling messages via gRPC and share intermediate results through a POSIX shared memory segment. Upon receiving a scheduling message from the dynamic DAG topology orchestration engine 2 via the gRPC protocol, each agent process loads its own role prompt and upstream context memory pointer, performs inference, and writes the output back to the shared memory segment.

[0030] The dynamic rerouting and logical circuit breaker module 4 subscribes to the output of each execution node on the current dynamic DAG topology and calculates the information entropy score and rule confidence. When a circuit breaker is triggered, it sends a suspension command and a remedial instantiation command to the dynamic DAG topology orchestration engine 2, and sends the circuit breaker and degradation snapshot to the Merkle tree white-box causal chain evidence storage module 5. Each execution node includes the corresponding nodes of macro analysis Agent 3a, factor calculation Agent 3b, and risk control compliance Agent 3c in the distributed Agent execution cluster 3, as well as the remedial Agent node dynamically instantiated after the circuit breaker.

[0031] Merkle tree white-box causal chain evidence module 5 collects the input hashes of each node, runs the Prompt, outputs the results and microsecond-level timestamps, constructs the Merkle tree and calculates the causal root hash, and writes it to the WORM medium.

[0032] The intelligent investment research and decision-making output terminal 6 renders a structured decision report with accompanying causal root hash values.

[0033] An example deployment environment of an embodiment of the system of the present invention is as follows: The system is deployed in a computing cluster consisting of 4 servers, including 1 master control server and 3 working computing servers; wherein, the master control server deploys a data access module 1, a dynamic DAG (initial dynamic directed acyclic graph) topology orchestration engine 2, a dynamic rerouting and logical circuit breaker module 4, and a Merkle tree white-box causal chain evidence storage module 5; the 3 working computing servers respectively deploy macro analysis agent 3a, factor calculation agent 3b, and risk control compliance agent 3c in the distributed agent execution cluster 3, and each agent process is independently bound to a GPU or a group of CPU cores of the corresponding working computing server; the intelligent investment research decision output terminal 6 runs on the investment research user interaction terminal; among the agent processes in the distributed agent execution cluster 3, and between the suspended agent process and the newly instantiated remedial agent process, intermediate inference results and context memory pointers are shared through a POSIX shared memory segment, and the shared memory segment locks physical pages with mlock to prevent page swapping; the WORM medium adopts an object storage bucket with legal hold enabled.

[0034] like Figures 2 to 6 As shown, an embodiment of the large-scale multi-agent collaboration and dynamic DAG orchestration method for intelligent FICC investment research of the present invention includes preliminary preparation steps and steps S101 to S105, and is presented as follows: Figure 3 The five-node investment research task shown is a comprehensive example: Node N1 Macroeconomic Analysis (in-degree 0), N2 Interest Rate Scenario Simulation (dependent on N1, in-degree 1), N3 Factor Calculation (dependent on N1 and N2, in-degree 2), N4 Risk Control and Compliance (dependent on N3, in-degree 1), and N5 Report Generation (dependent on N3 and N4, in-degree 2).

[0035] Preliminary steps: Connecting the market data middleware with the macroeconomic data interface to receive and preprocess FICC multi-source heterogeneous data streams. These FICC multi-source heterogeneous data streams include macroeconomic indicators, bond spreads, and tick-level transaction data. The data streams are structured and cleaned into a columnar structure with microsecond timestamps according to a unified model, and written into the context data base as the underlying context input for the DAG, available for reading by various agents. This step solves the technical problem of large language models lacking specialized financial domain context.

[0036] Step S101: Analyze the pre-defined logical dependency constraints of the multi-agent system, calculate the initial in-degree of each node, and construct the initial dynamic DAG topology. This step aims to solve the technical problem that complex investment research tasks are difficult to close using a single model. Directed edges are established based on the strong input-output dependencies between agents: for example, factor calculations must wait for macro-analysis results, so a directed edge N1→N3 is established. Specifically, after reading the dependency configuration, an adjacency list {N1:[N2,N3], N2:[N3], N3:[N4,N5],N4:[N5]} is generated, and the in-degree of each node is counted to obtain an in-degree list {N1:0, N2:1, N3:2, N4:1, N5:2}. N1, with an in-degree of 0, is pushed into the ready queue. This step parses the logical dependency constraints of the multi-agent system and constructs the initial dynamic DAG topology as follows: Figure 3 The initial dynamic directed acyclic graph (DAG) topology shown solves the technical problem that static DAGs cannot adaptively expand their topology based on abnormal intermediate results generated by inference.

[0037] Step S102: Schedule distributed multi-Agent nodes to perform investment research inference and calculation in parallel according to their in-degree priority. Combined with... Figure 3 Example of scheduling sequence: At time t0, the ready queue contains only N1, and the cluster allocates GPUs to trigger the execution of the macro-analysis agent; at time t1, N1 completes, and the dynamic DAG topology orchestration engine 2 decrements the in-degree of its direct downstream N2 and N3 by 1, making the in-degree table {N2:0, N3:1}. N2's in-degree becomes 0 and it is pushed into the ready queue for immediate execution, while N3 still needs to wait; at time t2, N2 completes, and N3's in-degree changes from 1 to 0 and is enqueued for execution; at time t3, after N3 completes, N4 is enqueued; at time t4, after N4 completes, N5 is enqueued. This scheduling with dynamic reduction of in-degree allows dependent nodes to run in parallel as much as possible, squeezing out concurrency performance and solving the technical problem of low efficiency in heterogeneous agent collaboration.

[0038] Step S103: Monitor the Agent node output in real time, calculate the information entropy score of the probability distribution and the confidence level of the logical verification rule, and trigger the dynamic circuit breaker mechanism. The dynamic rerouting and logical circuit breaker module 4 extracts the softmax probability distribution matrix when generating the Token sequence in real time, and calculates the information entropy according to H(X)=-∑p(x)log2p(x). Normal example: The factor calculation Agent output probability for the three candidates "over-allocated / average / under-allocated" is (0.4, 0.35, 0.25), then H=-(0.4×log2 0.4+0.35×log2 0.35+0.25×log2 0.25)≈1.55, which is lower than the entropy threshold of 1.8, so it is judged to be normal and allowed to pass. Anomaly Example: If the output of a node has a near-uniform probability distribution across the five candidate directions (0.2, 0.2, 0.2, 0.2, 0.2), then H = log2 5 ≈ 2.32 > 1.8, indicating extreme hesitation and high risk of illusion in the model, triggering an immediate circuit breaker. Simultaneously, the dynamic rerouting and logical circuit breaker module 4 performs confidence matching using the pre-configured business red line rule library from the expert system. For example, the rule "exposure to a single commodity shall not exceed 10% of net assets" has a calculated confidence level of 0.62, lower than the pre-set confidence threshold of 0.9, also triggering a circuit breaker. Circuit breaker intervention occurs if any indicator fails to meet the threshold. Regarding the calibration basis of the entropy threshold of 1.8: This embodiment stress-tests the output distribution of thousands of historical research reports, statistically finding that the 99th percentile of the information entropy of normal output is approximately 1.72, which is calibrated to 1.8 after reserving a safety margin. This step solves the technical problems of large model black boxes and logical illusions.

[0039] Step S104: Suspend the current node's original downstream node, instantiate the recovery Agent node, and reconstruct the computation graph by inheriting the context memory pointer. Combined with... Figure 4 and Figure 5 Assuming N3 triggers a circuit breaker, the dynamic DAG topology orchestration engine 2 sets N3 to the circuit breaker state, and its downstream nodes N4 and N5 to the suspended state, and dynamically instantiates the remediation verification Agent node N3'. At the memory safety level, the system reads the context memory block header of the suspended node N3 (including the memory starting address 0x7F3A0000, length 4KB, reference count RC, and read-write lock RW-Lock fields). Under read-write lock protection, the reference count RC is incremented from 1 to 2, and the memory starting address is directly mapped to the input buffer of the N3' process using mmap shared mapping, making both processes point to the same physical page, achieving zero-copy inheritance and avoiding the overhead of deep copying the entire context. The remediation execution process is as follows: N3' reads the upstream intermediate result in shared memory, changes the verification prompt, lowers the sampling temperature parameter from 0.7 to 0.2, regenerates it, and executes the entropy and rule verification in step S103 again; after successful verification, the dynamic DAG topology orchestration engine 2 reconnects the output of N3' to the original downstream node N4, such as... Figure 4As shown, the scheduling of the suspended branch is restored. Abnormal branch closed-loop strategy: If the remedial agent, after the set retry limit (3 times in this example), still outputs an entropy higher than the threshold, the system triggers a global degradation strategy: terminating the continued calculation of the DAG branch, throwing a "research logic blockage" exception upstream and labeling it as unreliable in the alarm log, and simultaneously sending the circuit breaker and degradation snapshot of the branch to the Merkle tree white-box causal chain evidence storage module 5 for evidence storage. Other unrelated node branches continue execution to prevent infinite loops from blocking the overall workflow. Actual testing shows that this dynamic circuit breaker expansion mechanism reduces peak memory pressure by 60% compared to the traditional deep copy scheme. This step solves the technical bottleneck of static workflows not being able to dynamically and adaptively expand, and addresses the technical problems of static DAGs not being able to dynamically expand and the high overhead of memory copying, through zero-copy inheritance of memory pointers.

[0040] Step S105: Extract a computation snapshot, construct a Merkle tree to calculate the causal root hash value, and output a structured decision conclusion. Combined with... Figure 6 The Merkle tree white-box causal chain evidence module 5 concatenates the input hash, execution prompt, output result, and microsecond-level timestamp of each node to calculate the SHA-256 hash, which is then used as the leaf node. That is, leaf L = SHA256(input hash || Prompt || output || microsecond timestamp). For example, if an execution generates five leaves L1, L2, L3, L4, and L5, since the number is odd, the last leaf L5 is automatically copied to make it even (denoted as L5'). Then, the hashes are concatenated pairwise: H12 = SHA256(L1 || L2), H34 = SHA256(L3 || L4), H55 = SHA256(L5 || L5'), and then upwards H1234 = SHA256(H12 || H34), finally resulting in the causal root hash = SHA256(H1234 || H55). The root hash is output along with the structured decision report and persisted to the WORM immutable medium. This embodiment enables legally reserved object storage buckets. Any tampering with historical node data will cause the root hash verification to fail, achieving 100% audit traceability of the investment research chain. This step solves the technical problem of the lack of causal chain tracing in decision-making and the difficulty of black-box architecture meeting the stringent compliance audit traceability requirements of financial institutions.

[0041] The present invention also provides a computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above method.

[0042] As can be seen from the above explanation of the method and system of the present invention: the present invention constructs an initial dynamic directed acyclic graph (DAG) by parsing the logical dependency constraints of multiple agents, and schedules distributed multiple agents to perform inference in parallel; extracts the information entropy and rule confidence of the agent output probability distribution in real time, executes a dynamic circuit breaker mechanism when a threshold is triggered, suspends downstream nodes and dynamically inserts a remedy agent node by inheriting the context memory pointer of the suspended node; finally, constructs the root hash value of the computation chain according to the Merkle tree structure and persists it.

[0043] This invention improves parallel efficiency through an in-degree decreasing scheduling mechanism, specifically intercepts logical illusions through a large model information entropy circuit breaker mechanism, achieves zero-copy imputation and anomaly closure for remedial agents through a context memory pointer zero-copy inheritance mechanism, and finally constructs the execution snapshot as a Merkle tree for evidence storage, ensuring 100% tamper-proof computation and eliminating compliance audit blind spots in intelligent investment research. This invention combines four technical means to overcome the black-box illusion problem of traditional large models in financial scenarios, and solves the technical bottleneck of static workflows not being able to dynamically and adaptively expand through memory pointer zero-copy inheritance, achieving highly reliable intelligent investment research decision-making with white-box traceability.

[0044] The above embodiments are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Although embodiments of the present invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for large-scale multi-agent collaboration and dynamic DAG orchestration in intelligent investment research for FICC, characterized in that, It includes the following steps: Step S101: parse the logical dependency constraints between multiple agents, calculate the initial in-degree of the nodes, and construct the initial dynamic directed acyclic graph topology. Step S102: Schedule distributed multi-agent nodes to execute financial investment research reasoning and calculation tasks in parallel according to the in-degree priority of the dynamic directed acyclic graph topology; Step S103: Monitor the output data of each intelligent agent node in real time, calculate the information entropy score of the output probability distribution of the intelligent agent node, and calculate the confidence level by matching the preset logical verification rules. If the information entropy score is higher than the set entropy threshold or the confidence level is lower than the preset confidence threshold, trigger the dynamic circuit breaker mechanism. Step S104: The dynamic circuit breaker mechanism suspends the original downstream node of the current agent node, instantiates a remedy agent node, and the remedy agent node inherits the context memory pointer of the suspended node to reconstruct the memory allocation of the computation graph. Step S105: Extract the input hash of each agent node, run the Prompt, output the result and microsecond-level timestamp, construct the Merkle tree structure and calculate the global causal root hash value, and output the structured decision conclusion with the causal root hash value.

2. The method for large-scale multi-agent collaboration and dynamic DAG orchestration of intelligent investment research for FICC, as described in claim 1, is characterized in that... Before step S101, the method further includes: receiving a multi-source heterogeneous FICC data stream containing macroeconomic indicators, bond spreads, and tick-level transaction data from the market data middleware and macro data interface, and structuring and cleaning it into a columnar structure with microsecond timestamps according to a unified mode, which serves as the underlying context input data of the dynamic directed acyclic graph topology.

3. The method for large-scale multi-agent collaboration and dynamic DAG orchestration of intelligent investment research for FICC, as described in claim 1, is characterized in that... Step S101 includes maintaining a node table and an adjacency table in memory; each record in the node table contains a node number, Agent type, in-degree counter, and state machine field, and the state machine values ​​include five types: waiting, running, suspended, reviewed, and circuit breaker; the adjacency table records a list of direct downstream node numbers for each node; the ready queue is implemented as a lock-free first-in-first-out queue.

4. The method for large-scale multi-agent collaboration and dynamic DAG orchestration of intelligent investment research for FICC, as described in claim 3, is characterized in that... The process of maintaining the node table and adjacency table in memory includes establishing directed edges based on the strong input-output dependencies between agents: after reading the dependency configuration, an adjacency table is generated, the in-degree of each node is counted to obtain the in-degree table, and N1 with an in-degree of 0 is pushed into the ready queue.

5. The method for large-scale multi-agent collaboration and dynamic DAG arrangement of intelligent investment research for FICC as described in claim 1, characterized in that, The specific scheduling process in step S102 includes: Push the agent node with an in-degree of 0 into the ready queue, allocate GPU / CPU computing resources, and trigger execution; After the current agent node finishes execution, the in-degree of its direct downstream node is decremented by 1; if the in-degree of the downstream node decreases to 0, it is pushed back into the ready queue for scheduling.

6. The method for large-scale multi-agent collaboration and dynamic DAG orchestration of intelligent investment research for FICC, as described in claim 1, is characterized in that... The specific process of inheriting the context memory pointer of the suspended node in step S104 includes: Read the memory starting address and context state handle of the suspended node, and perform concurrency protection through read-write locks and reference counting mechanisms; The memory starting address is directly mapped to the input buffer of the newly instantiated remedial agent node, achieving zero-copy memory sharing.

7. The method for large-scale multi-agent collaboration and dynamic DAG orchestration of intelligent investment research for FICC, as described in claim 6, is characterized in that... The remedial execution process is as follows: the remedial verification agent node N3' reads the upstream intermediate result in the shared memory, replaces the verification prompt and lowers the sampling temperature parameter before regenerating, and executes step S103 again; after the verification is passed, the output of N3' is reconnected to the original downstream N4 to restore the scheduling of the suspended branch.

8. The method for large-scale multi-agent collaboration and dynamic DAG orchestration of intelligent investment research for FICC as described in claim 1, characterized in that, The dynamic circuit breaker mechanism also includes an abnormal retry closed-loop strategy: if the output information entropy of the remedial agent node is still higher than the threshold after a limited number of retries, a system-level degradation strategy is triggered to terminate the graph computation of the computation branch in advance and mark the alarm log.

9. A large-scale multi-agent collaboration and dynamic DAG orchestration system for intelligent investment research in FICC, implemented based on the method described in any one of claims 1-8, characterized in that, Include: The data input module is used to receive FICC multi-source data streams; A dynamic DAG orchestration engine is used to resolve agent dependencies and maintain dynamic topology; The distributed intelligent agent execution cluster includes macro-analysis intelligent agents, factor calculation intelligent agents, and risk control and compliance intelligent agents running in parallel; The dynamic rerouting and logical circuit breaker module is used to calculate the information entropy score and schedule the remedial agent node and memory pointer inheritance when the circuit breaker is triggered; Merkle tree white-box causal chain persistence module, used to generate and store Merkle root hash values; The decision output terminal is used to display the conclusions with causal root hash values.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 8.

Images