Programmable Data Plane Telemetry for AI Network Operations
MAR 17, 20269 MIN READ
Generate Your Research Report Instantly with AI Agent
Patsnap Eureka helps you evaluate technical feasibility & market potential.
Programmable Data Plane Telemetry Background and AI Network Goals
Programmable data plane telemetry represents a paradigm shift in network monitoring and management, evolving from traditional static monitoring approaches to dynamic, software-defined telemetry systems. This technology emerged from the limitations of conventional network monitoring tools that relied on fixed protocols like SNMP and NetFlow, which provided limited visibility into network behavior and insufficient granularity for modern network operations.
The evolution of data plane telemetry has been driven by the increasing complexity of network infrastructures and the growing demand for real-time network insights. Traditional monitoring approaches struggled with scalability issues, high overhead, and inflexibility in adapting to diverse network environments. The introduction of programmable switches and software-defined networking (SDN) principles laid the foundation for more sophisticated telemetry capabilities.
Modern programmable data plane telemetry leverages technologies such as P4 (Programming Protocol-independent Packet Processors) and In-band Network Telemetry (INT) to enable fine-grained, customizable network monitoring. These technologies allow network operators to define custom telemetry collection points, specify which network metrics to gather, and determine how data should be processed and exported.
The integration of artificial intelligence into network operations has created unprecedented demands for comprehensive, real-time network data. AI-driven network management systems require continuous streams of high-quality telemetry data to make intelligent decisions about traffic optimization, anomaly detection, predictive maintenance, and automated remediation. This necessity has accelerated the development of programmable telemetry solutions that can adapt to AI algorithms' specific data requirements.
The primary technical objectives of programmable data plane telemetry for AI network operations include achieving microsecond-level latency measurements, enabling per-packet visibility without significant performance degradation, and providing programmable data collection interfaces that can be dynamically reconfigured based on AI model requirements. These systems aim to deliver comprehensive network state information while maintaining minimal impact on network performance and scalability.
Contemporary implementations focus on creating telemetry frameworks that can seamlessly integrate with machine learning pipelines, supporting both batch processing for historical analysis and real-time streaming for immediate decision-making. The technology targets the elimination of blind spots in network visibility while providing the flexibility to adapt monitoring strategies as AI algorithms evolve and network requirements change.
The evolution of data plane telemetry has been driven by the increasing complexity of network infrastructures and the growing demand for real-time network insights. Traditional monitoring approaches struggled with scalability issues, high overhead, and inflexibility in adapting to diverse network environments. The introduction of programmable switches and software-defined networking (SDN) principles laid the foundation for more sophisticated telemetry capabilities.
Modern programmable data plane telemetry leverages technologies such as P4 (Programming Protocol-independent Packet Processors) and In-band Network Telemetry (INT) to enable fine-grained, customizable network monitoring. These technologies allow network operators to define custom telemetry collection points, specify which network metrics to gather, and determine how data should be processed and exported.
The integration of artificial intelligence into network operations has created unprecedented demands for comprehensive, real-time network data. AI-driven network management systems require continuous streams of high-quality telemetry data to make intelligent decisions about traffic optimization, anomaly detection, predictive maintenance, and automated remediation. This necessity has accelerated the development of programmable telemetry solutions that can adapt to AI algorithms' specific data requirements.
The primary technical objectives of programmable data plane telemetry for AI network operations include achieving microsecond-level latency measurements, enabling per-packet visibility without significant performance degradation, and providing programmable data collection interfaces that can be dynamically reconfigured based on AI model requirements. These systems aim to deliver comprehensive network state information while maintaining minimal impact on network performance and scalability.
Contemporary implementations focus on creating telemetry frameworks that can seamlessly integrate with machine learning pipelines, supporting both batch processing for historical analysis and real-time streaming for immediate decision-making. The technology targets the elimination of blind spots in network visibility while providing the flexibility to adapt monitoring strategies as AI algorithms evolve and network requirements change.
Market Demand for AI-Driven Network Operations and Telemetry
The global network operations market is experiencing unprecedented transformation driven by the exponential growth of data traffic, cloud adoption, and digital transformation initiatives across industries. Traditional network monitoring and management approaches are increasingly inadequate for handling the complexity and scale of modern distributed systems, creating substantial demand for AI-driven solutions that can provide real-time insights and automated decision-making capabilities.
Enterprise networks are becoming increasingly complex with the proliferation of multi-cloud environments, edge computing deployments, and IoT devices. Organizations require sophisticated telemetry systems that can collect, process, and analyze network data at unprecedented scales while maintaining low latency and high accuracy. The demand for programmable data plane telemetry specifically stems from the need for granular, real-time visibility into network behavior that traditional SNMP-based monitoring cannot provide.
Service providers and cloud operators face mounting pressure to deliver consistent performance while managing costs and ensuring security across vast network infrastructures. The ability to leverage AI for predictive analytics, anomaly detection, and automated remediation has become a critical competitive differentiator. This drives significant investment in telemetry platforms that can feed high-quality, structured data to machine learning algorithms for network optimization.
The telecommunications industry's transition to 5G networks and network function virtualization creates additional demand for intelligent network operations. These technologies require dynamic resource allocation, real-time performance optimization, and predictive maintenance capabilities that can only be achieved through AI-driven approaches supported by comprehensive telemetry data.
Financial services, healthcare, and other regulated industries are particularly driving demand for AI-powered network security and compliance monitoring. These sectors require continuous visibility into network traffic patterns, user behavior analytics, and threat detection capabilities that rely heavily on programmable telemetry systems capable of adapting to evolving security requirements.
The emergence of intent-based networking and self-healing network architectures represents a significant market opportunity. Organizations seek solutions that can translate business policies into network configurations automatically while continuously monitoring and adjusting performance based on real-time telemetry data processed through AI algorithms.
Enterprise networks are becoming increasingly complex with the proliferation of multi-cloud environments, edge computing deployments, and IoT devices. Organizations require sophisticated telemetry systems that can collect, process, and analyze network data at unprecedented scales while maintaining low latency and high accuracy. The demand for programmable data plane telemetry specifically stems from the need for granular, real-time visibility into network behavior that traditional SNMP-based monitoring cannot provide.
Service providers and cloud operators face mounting pressure to deliver consistent performance while managing costs and ensuring security across vast network infrastructures. The ability to leverage AI for predictive analytics, anomaly detection, and automated remediation has become a critical competitive differentiator. This drives significant investment in telemetry platforms that can feed high-quality, structured data to machine learning algorithms for network optimization.
The telecommunications industry's transition to 5G networks and network function virtualization creates additional demand for intelligent network operations. These technologies require dynamic resource allocation, real-time performance optimization, and predictive maintenance capabilities that can only be achieved through AI-driven approaches supported by comprehensive telemetry data.
Financial services, healthcare, and other regulated industries are particularly driving demand for AI-powered network security and compliance monitoring. These sectors require continuous visibility into network traffic patterns, user behavior analytics, and threat detection capabilities that rely heavily on programmable telemetry systems capable of adapting to evolving security requirements.
The emergence of intent-based networking and self-healing network architectures represents a significant market opportunity. Organizations seek solutions that can translate business policies into network configurations automatically while continuously monitoring and adjusting performance based on real-time telemetry data processed through AI algorithms.
Current State of Programmable Data Plane Telemetry Technologies
Programmable data plane telemetry has emerged as a critical technology for modern network infrastructure, particularly in environments requiring real-time monitoring and adaptive network management. The current landscape is dominated by several key technological approaches, each offering distinct capabilities for network visibility and control.
P4-based telemetry solutions represent the most mature segment of programmable data plane technologies. Major implementations include In-band Network Telemetry (INT) and Postcard-based Telemetry (PBT), which enable packet-level monitoring with microsecond granularity. These solutions leverage programmable switches from vendors like Barefoot Networks (now Intel), Broadcom, and Mellanox to embed telemetry metadata directly into network packets or generate separate telemetry streams.
eBPF (Extended Berkeley Packet Filter) has gained significant traction as a kernel-level programmable telemetry solution. Current implementations allow for high-performance packet processing and telemetry collection at the host level, with frameworks like Cilium and Katran demonstrating production-ready capabilities. eBPF-based solutions excel in container and cloud-native environments, providing detailed application-level network insights.
OpenFlow-based telemetry, while considered legacy compared to newer approaches, continues to play a role in hybrid deployments. Software-defined networking controllers can program flow tables to mirror or sample traffic for telemetry purposes, though with limited granularity compared to P4-based solutions.
Current commercial platforms demonstrate varying levels of maturity. Cisco's Application Centric Infrastructure (ACI) integrates programmable telemetry with policy enforcement, while Arista's CloudVision platform combines streaming telemetry with machine learning analytics. Open-source initiatives like ONOS and OpenDaylight provide controller frameworks supporting programmable telemetry applications.
The integration challenges remain significant, particularly regarding standardization across different hardware platforms and vendor ecosystems. Current solutions often require specialized hardware or specific kernel versions, limiting widespread adoption. Performance overhead and scalability concerns persist, especially when implementing comprehensive telemetry in high-throughput production networks.
Emerging standards like P4Runtime and gNMI (gRPC Network Management Interface) are beginning to address interoperability issues, enabling more consistent telemetry implementations across diverse network infrastructures. However, the technology landscape remains fragmented, with proprietary solutions often outperforming standardized approaches in specific use cases.
P4-based telemetry solutions represent the most mature segment of programmable data plane technologies. Major implementations include In-band Network Telemetry (INT) and Postcard-based Telemetry (PBT), which enable packet-level monitoring with microsecond granularity. These solutions leverage programmable switches from vendors like Barefoot Networks (now Intel), Broadcom, and Mellanox to embed telemetry metadata directly into network packets or generate separate telemetry streams.
eBPF (Extended Berkeley Packet Filter) has gained significant traction as a kernel-level programmable telemetry solution. Current implementations allow for high-performance packet processing and telemetry collection at the host level, with frameworks like Cilium and Katran demonstrating production-ready capabilities. eBPF-based solutions excel in container and cloud-native environments, providing detailed application-level network insights.
OpenFlow-based telemetry, while considered legacy compared to newer approaches, continues to play a role in hybrid deployments. Software-defined networking controllers can program flow tables to mirror or sample traffic for telemetry purposes, though with limited granularity compared to P4-based solutions.
Current commercial platforms demonstrate varying levels of maturity. Cisco's Application Centric Infrastructure (ACI) integrates programmable telemetry with policy enforcement, while Arista's CloudVision platform combines streaming telemetry with machine learning analytics. Open-source initiatives like ONOS and OpenDaylight provide controller frameworks supporting programmable telemetry applications.
The integration challenges remain significant, particularly regarding standardization across different hardware platforms and vendor ecosystems. Current solutions often require specialized hardware or specific kernel versions, limiting widespread adoption. Performance overhead and scalability concerns persist, especially when implementing comprehensive telemetry in high-throughput production networks.
Emerging standards like P4Runtime and gNMI (gRPC Network Management Interface) are beginning to address interoperability issues, enabling more consistent telemetry implementations across diverse network infrastructures. However, the technology landscape remains fragmented, with proprietary solutions often outperforming standardized approaches in specific use cases.
Existing Programmable Telemetry Solutions for AI Networks
01 In-band telemetry and packet header processing
Programmable data plane telemetry can be implemented through in-band telemetry mechanisms where telemetry data is embedded directly into packet headers. This approach allows network devices to collect and process telemetry information as packets traverse the network, enabling real-time monitoring of network conditions, latency, and path information. The programmable data plane can be configured to insert, extract, and process telemetry metadata at line rate without impacting packet forwarding performance.- In-band telemetry and packet header processing: Programmable data plane telemetry can be implemented through in-band telemetry mechanisms where telemetry data is embedded directly into packet headers. This approach allows network devices to collect and process telemetry information as packets traverse the network, enabling real-time monitoring of network conditions. The programmable data plane can be configured to insert, extract, and process telemetry metadata within packet headers, providing visibility into packet paths, latency, and device states without requiring separate control plane communications.
- Programmable pipeline architecture for telemetry collection: Telemetry systems utilize programmable pipeline architectures that enable flexible configuration of data collection points throughout the packet processing stages. These architectures allow network operators to define custom telemetry collection logic, specify which metrics to gather, and determine how data should be aggregated and exported. The programmable nature enables adaptation to different network requirements and protocols while maintaining high-speed packet forwarding performance.
- Telemetry data aggregation and export mechanisms: Advanced telemetry systems incorporate mechanisms for aggregating collected data and exporting it to external monitoring systems. These mechanisms include buffering strategies, sampling techniques, and compression methods to manage the volume of telemetry data. The programmable data plane can be configured to filter, summarize, and format telemetry information before transmission to collectors, ensuring efficient use of network bandwidth while providing comprehensive visibility into network operations.
- Programmable telemetry for network troubleshooting and diagnostics: Programmable data plane telemetry enables sophisticated network troubleshooting and diagnostic capabilities by allowing operators to dynamically configure monitoring parameters based on specific issues or conditions. The system can be programmed to track packet flows, identify anomalies, measure performance metrics, and correlate events across multiple network devices. This flexibility supports rapid problem identification and resolution by providing detailed insights into network behavior at the data plane level.
- Integration with network management and analytics platforms: Telemetry systems are designed to integrate with broader network management and analytics platforms, enabling centralized monitoring and analysis of network-wide telemetry data. The programmable data plane provides standardized interfaces and protocols for communicating telemetry information to external systems, supporting automated network optimization, capacity planning, and security monitoring. This integration allows for correlation of data plane telemetry with control plane information and application-level metrics to provide comprehensive network visibility.
02 Programmable pipeline architecture for telemetry collection
Telemetry systems utilize programmable pipeline architectures that enable flexible configuration of data collection points throughout the packet processing stages. These architectures allow operators to define custom telemetry collection logic, specify which metrics to gather, and determine how telemetry data should be aggregated and exported. The programmable nature enables adaptation to different monitoring requirements and network conditions without hardware modifications.Expand Specific Solutions03 Telemetry data aggregation and export mechanisms
Advanced mechanisms for aggregating and exporting telemetry data from programmable data planes enable efficient collection of network statistics and performance metrics. These systems implement intelligent sampling, filtering, and compression techniques to manage the volume of telemetry data while maintaining accuracy. Export protocols and formats are designed to integrate with network management systems and analytics platforms for comprehensive visibility.Expand Specific Solutions04 Real-time network monitoring and analytics
Programmable data plane telemetry enables real-time network monitoring capabilities by collecting granular performance metrics including latency, jitter, packet loss, and queue depths. The telemetry infrastructure supports continuous monitoring of network health and can trigger alerts or automated responses based on predefined thresholds. This real-time visibility facilitates rapid troubleshooting and proactive network management.Expand Specific Solutions05 Programmable telemetry for traffic engineering and optimization
Telemetry data collected from programmable data planes can be utilized for traffic engineering and network optimization purposes. By analyzing flow-level statistics, path characteristics, and resource utilization metrics, network operators can make informed decisions about routing, load balancing, and capacity planning. The programmable nature allows customization of telemetry collection to support specific optimization algorithms and traffic management strategies.Expand Specific Solutions
Key Players in Programmable Network and AI Operations Industry
The programmable data plane telemetry for AI network operations represents an emerging technology sector in the early growth stage, driven by the increasing demand for AI-driven network automation and real-time visibility. The market is experiencing rapid expansion as enterprises seek intelligent network monitoring solutions to support complex AI workloads. Technology maturity varies significantly across players, with established networking giants like Huawei Technologies, Juniper Networks, and Mellanox Technologies leading in infrastructure capabilities, while specialized companies such as Aviz Networks and NIKSUN focus on AI-driven network observability and packet capture solutions. Academic institutions including Beijing Jiaotong University and Beihang University contribute foundational research, while telecommunications providers like China Telecom's Technology Innovation Center drive practical implementations. The competitive landscape shows a convergence of traditional networking hardware vendors, emerging software-defined networking specialists, and research organizations, indicating a technology transition phase where programmable telemetry capabilities are becoming critical for next-generation AI network operations.
Huawei Technologies Co., Ltd.
Technical Solution: Huawei has developed comprehensive programmable data plane telemetry solutions integrated with their CloudFabric network architecture. Their approach leverages P4-programmable switches with In-band Network Telemetry (INT) capabilities, enabling real-time collection of network state information including latency, jitter, queue depth, and packet loss metrics. The solution incorporates AI-driven analytics engines that process telemetry data streams in real-time to detect anomalies, predict network congestion, and automatically trigger remediation actions. Their telemetry framework supports both hop-by-hop and end-to-end monitoring with microsecond-level granularity, utilizing custom ASICs optimized for high-speed packet processing and metadata insertion without impacting forwarding performance.
Strengths: Comprehensive end-to-end solution with proven deployment scale, strong integration with AI analytics. Weaknesses: Proprietary implementation may limit interoperability with third-party network equipment.
Mellanox Technologies Ltd.
Technical Solution: Mellanox (now part of NVIDIA) provides programmable data plane telemetry through their ConnectX SmartNIC and Spectrum switch platforms. Their solution implements hardware-accelerated telemetry collection using RDMA over Converged Ethernet (RoCE) and InfiniBand protocols optimized for AI/ML workloads. The system features programmable packet processing engines that can insert telemetry metadata directly in the data path with sub-microsecond latency overhead. Their telemetry framework includes specialized counters for GPU-to-GPU communication patterns, collective operations monitoring, and real-time congestion detection. The solution integrates with NVIDIA's AI Enterprise software stack to provide closed-loop optimization for distributed training workloads, automatically adjusting network parameters based on telemetry insights.
Strengths: Hardware-accelerated performance with minimal latency impact, excellent integration with AI/GPU workloads. Weaknesses: Primarily focused on NVIDIA ecosystem, limited programmability compared to pure P4 solutions.
Core Innovations in Data Plane Programming for AI Telemetry
Method for Dynamic Resource Scheduling of Programmable Dataplanes for Network Telemetry
PatentActiveUS20230161769A1
Innovation
- The Dynamic Approximate Telemetry Operation Scheduler (DynATOS) reframes telemetry systems as resource schedulers, using a reconfigurable approach to dynamically schedule and execute telemetry queries on a programmable dataplane device, balancing accuracy and latency while optimizing resource usage through time-division approximation and multi-objective optimization.
End-to-end RDMA telemetry system
PatentActiveUS11876691B2
Innovation
- An end-to-end RDMA telemetry system comprising distributed programmable data planes that extract network-level information and local RDMA tracers that identify host-level operations, with a telemetry collector generating reports for real-time monitoring at the RDMA protocol level across all RDMA-enabled workloads.
Network Security and Privacy Considerations for AI Telemetry
The integration of programmable data plane telemetry with AI-driven network operations introduces significant security and privacy challenges that must be carefully addressed to ensure robust network infrastructure protection. As telemetry systems collect vast amounts of granular network data for AI analysis, they create new attack vectors and privacy exposure points that traditional network security frameworks may not adequately cover.
Data collection security represents a primary concern, as programmable data planes generate high-volume, real-time telemetry streams containing sensitive network topology information, traffic patterns, and performance metrics. These data flows require end-to-end encryption and secure transmission protocols to prevent interception and manipulation by malicious actors. The programmable nature of data planes also introduces risks of telemetry poisoning attacks, where adversaries could inject false or misleading data to compromise AI decision-making processes.
Privacy preservation becomes particularly complex when telemetry data contains user behavior patterns and application-specific information. Organizations must implement data anonymization and differential privacy techniques to protect individual user privacy while maintaining the statistical utility required for effective AI operations. The challenge lies in balancing data granularity needed for accurate AI insights with privacy protection requirements mandated by regulatory frameworks.
Access control and authentication mechanisms must be strengthened to secure the interfaces between programmable data planes and AI processing systems. Multi-factor authentication, role-based access controls, and secure API gateways are essential to prevent unauthorized access to telemetry configuration and data streams. Additionally, the distributed nature of modern networks requires consistent security policy enforcement across multiple data plane instances.
AI model security presents another critical dimension, as adversaries could exploit telemetry data to reverse-engineer network configurations or launch adversarial attacks against machine learning models. Implementing secure model training environments, federated learning approaches, and robust model validation processes helps mitigate these risks while preserving the effectiveness of AI-driven network operations.
Data collection security represents a primary concern, as programmable data planes generate high-volume, real-time telemetry streams containing sensitive network topology information, traffic patterns, and performance metrics. These data flows require end-to-end encryption and secure transmission protocols to prevent interception and manipulation by malicious actors. The programmable nature of data planes also introduces risks of telemetry poisoning attacks, where adversaries could inject false or misleading data to compromise AI decision-making processes.
Privacy preservation becomes particularly complex when telemetry data contains user behavior patterns and application-specific information. Organizations must implement data anonymization and differential privacy techniques to protect individual user privacy while maintaining the statistical utility required for effective AI operations. The challenge lies in balancing data granularity needed for accurate AI insights with privacy protection requirements mandated by regulatory frameworks.
Access control and authentication mechanisms must be strengthened to secure the interfaces between programmable data planes and AI processing systems. Multi-factor authentication, role-based access controls, and secure API gateways are essential to prevent unauthorized access to telemetry configuration and data streams. Additionally, the distributed nature of modern networks requires consistent security policy enforcement across multiple data plane instances.
AI model security presents another critical dimension, as adversaries could exploit telemetry data to reverse-engineer network configurations or launch adversarial attacks against machine learning models. Implementing secure model training environments, federated learning approaches, and robust model validation processes helps mitigate these risks while preserving the effectiveness of AI-driven network operations.
Standardization Efforts in Programmable Network Telemetry
The standardization of programmable network telemetry has emerged as a critical initiative to ensure interoperability and consistency across diverse network infrastructures supporting AI operations. Multiple industry organizations and standards bodies are actively developing frameworks to address the unique requirements of AI-driven network environments.
The P4 Language Consortium has been instrumental in advancing programmable data plane standards through the P4 specification, which enables network operators to define custom telemetry collection mechanisms directly within network hardware. The P4Runtime API provides standardized interfaces for controlling programmable switches and collecting telemetry data, forming a foundation for AI network operations that require real-time visibility into packet processing behaviors.
The Open Networking Foundation (ONF) has contributed significantly through the Stratum project, which defines a vendor-neutral switch operating system that supports standardized telemetry interfaces. This initiative enables consistent telemetry data collection across heterogeneous network equipment, essential for AI systems that require normalized data inputs for effective network optimization and anomaly detection.
Industry collaboration through the Broadband Forum has resulted in the development of TR-383 specification, which addresses common data models for network function virtualization telemetry. This standard provides structured approaches for collecting and reporting performance metrics from virtualized network functions, supporting AI applications that monitor and optimize network service delivery.
The Internet Engineering Task Force (IETF) has established working groups focused on network telemetry standardization, including efforts on YANG data models for streaming telemetry and gRPC-based network management protocols. These standards facilitate the integration of programmable telemetry systems with existing network management frameworks while supporting the high-frequency data collection requirements of AI-driven network operations.
Recent standardization efforts have emphasized the development of common telemetry data formats and APIs that enable seamless integration between different vendor solutions. The emergence of OpenConfig as a collaborative effort between network operators and equipment vendors has produced standardized YANG models specifically designed for streaming telemetry applications, providing consistent data structures that AI systems can reliably process across multi-vendor environments.
The P4 Language Consortium has been instrumental in advancing programmable data plane standards through the P4 specification, which enables network operators to define custom telemetry collection mechanisms directly within network hardware. The P4Runtime API provides standardized interfaces for controlling programmable switches and collecting telemetry data, forming a foundation for AI network operations that require real-time visibility into packet processing behaviors.
The Open Networking Foundation (ONF) has contributed significantly through the Stratum project, which defines a vendor-neutral switch operating system that supports standardized telemetry interfaces. This initiative enables consistent telemetry data collection across heterogeneous network equipment, essential for AI systems that require normalized data inputs for effective network optimization and anomaly detection.
Industry collaboration through the Broadband Forum has resulted in the development of TR-383 specification, which addresses common data models for network function virtualization telemetry. This standard provides structured approaches for collecting and reporting performance metrics from virtualized network functions, supporting AI applications that monitor and optimize network service delivery.
The Internet Engineering Task Force (IETF) has established working groups focused on network telemetry standardization, including efforts on YANG data models for streaming telemetry and gRPC-based network management protocols. These standards facilitate the integration of programmable telemetry systems with existing network management frameworks while supporting the high-frequency data collection requirements of AI-driven network operations.
Recent standardization efforts have emphasized the development of common telemetry data formats and APIs that enable seamless integration between different vendor solutions. The emergence of OpenConfig as a collaborative effort between network operators and equipment vendors has produced standardized YANG models specifically designed for streaming telemetry applications, providing consistent data structures that AI systems can reliably process across multi-vendor environments.
Unlock deeper insights with Patsnap Eureka Quick Research — get a full tech report to explore trends and direct your research. Try now!
Generate Your Research Report Instantly with AI Agent
Supercharge your innovation with Patsnap Eureka AI Agent Platform!