OpenDerisk: AI-SRE Diagnostic Framework

• OpenDerisk is an open-source multi-agent framework designed specifically for diagnostic SRE, focusing on investigative root cause analysis in distributed, cloud-native systems.
• It employs a modular architecture incorporating a pluggable reasoning engine, Knowledge Engine, and standardized Model Context Protocol for effective multi-agent collaboration.
• Industrial evaluations at Ant Group show its robustness with over 3,000 daily active users and 60,000 diagnostic runs per day, despite increased latency due to complex agent coordination.

OpenDerisk is an industrial, open-source multi-agent framework for Site Reliability Engineering that is explicitly designed around diagnostic work rather than generic task automation. It is presented as a specialized AI-SRE system for root cause analysis, trace analysis, risk assessment, and related operational investigations in distributed, microservice-based, cloud-native production systems. Its defining components are a diagnostic-native collaboration model, a pluggable reasoning engine, a Knowledge Engine, and a standardized Model Context Protocol; in the reported production deployment at Ant Group, it served over 3,000 daily active users and executed more than 60,000 diagnostic runs per day (Di et al., 15 Oct 2025).

1. Definition and investigative orientation

OpenDerisk is framed around a specific claim about SRE work: incident handling is fundamentally investigative, causal, and hypothesis-driven rather than a mere pattern-matching or task-execution problem. In the paper’s formulation, an engineer performing root cause analysis must combine heterogeneous signals such as logs, metrics, traces, alarms, code changes, and environment changes; form hypotheses about causal chains; test those hypotheses by gathering evidence; and use domain and system knowledge to explain why an incident occurred. The framework therefore targets workflows such as RCA, trace analysis, and risk assessment rather than general software automation (Di et al., 15 Oct 2025).

This positioning is developed against two classes of prior systems. Traditional AI and AIOps methods, including causal discovery, dependency graph analysis, anomaly detection, and graph-based RCA, are described as useful for locating abnormalities but limited in deep semantic explanation. General-purpose LLM and multi-agent systems, including MetaGPT, TaskWeaver, AutoGen-style frameworks, SWE-agent, OpenDevin, and OpenRCA, are treated as insufficiently specialized because they are primarily optimized for generative software tasks or rely on a generalist problem-solving model rather than deep multi-domain diagnostic coordination. OpenDerisk is thus defined not as a generic agent shell, but as a diagnostic-native framework intended to emulate how expert SRE teams investigate incidents.

A common misconception is that OpenDerisk is simply another orchestration wrapper around LLMs. The paper argues for a narrower identity: the framework is specialized for causal diagnosis in production systems, where long, evolving context, heterogeneous operational evidence, and domain-specific reasoning are first-class requirements rather than secondary implementation details.

2. System architecture and execution flow

The architecture is organized as an end-to-end diagnostic pipeline with three principal stages: a Perception Layer, the DeRisk System as the core decision and execution engine, and an Analysis and Reporting Layer, with Human-in-the-Loop feedback surrounding the pipeline. The Perception Layer ingests incident indicators including Log Alarms, anomalous App Behavior, Environment Changes, and code changes from GitHub. These inputs constitute the initial state for investigation (Di et al., 15 Oct 2025).

Within the DeRisk System, execution is coordinated through a multi-agent subsystem, a reasoning engine, a Knowledge Engine, MCP-mediated tool access, and orchestration logic. Collaboration can switch dynamically among Single-Agent, TeamMode, and GroupMode. The paper characterizes these operationally rather than algorithmically: Single-Agent is used for simpler scenarios, TeamMode for a coordinated bounded task, and GroupMode for broader collaboration with richer shared context. A memory module supports these collaboration levels.

The clearest control loop is the V3 pattern described in the evaluation sections. A central Supervisor agent decomposes the task. Deterministic subproblems are dispatched via ToolCall to a Workflow Engine, whereas complex subproblems are sent by Handoff to specialized Sub-Agents. Specialists gather evidence and return structured summaries; the Supervisor then synthesizes findings and the reporting layer produces Diagnostic Reports, Root Cause Location, and Handling opinions. This divide-and-conquer structure is the operational core of the framework.

The architecture is also shaped by three explicit optimization objectives stated in the paper: Multi-Agent Collaboration Optimization, Context Engineering Optimization, and System-Level Reinforcement Learning. The design rejects rigid Workflow systems as too inflexible and rejects current Agentic RL as a primary basis because it is still immature and expensive. Instead, OpenDerisk is built on a Multi-Agent ReAct paradigm, presented as the most practical balance of flexibility, intelligence, and engineering tractability.

3. Agents, reasoning modes, knowledge grounding, and context engineering

OpenDerisk uses a specialist-based agent model. The default “digital employees” listed in the paper are SRE-Agent, Code-Agent, Data-Agent, Vis-Agent, and ReportAgent, and the text also mentions examples such as OS-Agent, Config-Agent, and scenario-specific agents. Each specialized agent inherits from a base Agent abstraction that standardizes a Profile, a set of Tools, a Planning engine, and a Memory module. The collaboration pattern mirrors an incident team: a coordinator frames and decomposes the problem, specialists investigate domain-specific subproblems, and structured findings are combined into a causal narrative (Di et al., 15 Oct 2025).

The reasoning engine is pluggable and supports three modes. Dynamic LLM ReAct Mode is used for exploratory diagnosis; Deterministic SOP Mode is used for routine tasks that should follow fixed procedures; and RL Dynamic Mode is intended for self-optimization. The paper does not provide a formal switching policy or mathematical planner, but it repeatedly describes the reasoning loop as ReAct-style iterative sense-making: reason, invoke tools, observe outputs, refine the line of inquiry, escalate to specialists when needed, and synthesize evidence.

Knowledge grounding is handled by the K-Engine, which transforms raw enterprise knowledge into a retrievable substrate through five stages: Knowledge Data Parsing, Knowledge Data Chunking, Semantic Enrichment / Augmentation, Hybrid Indexing, and Knowledge Active Learning / Update. The indexing layer combines KV Index, Vector Index, Full-text Index, and Knowledge Graph Index. The enrichment stage extracts knowledge, FAQ pairs, and domain entity relations. The update stage is intended to preserve knowledge freshness and enable knowledge correction. In the framework’s logic, this addresses the gap between pretrained model knowledge and proprietary, rapidly changing enterprise operational knowledge.

Tool and service integration is standardized through MCP. The protocol is described as a formal contract specifying how tools are described and used by agents, thereby enabling modularity, extensibility, and standardized integration. The paper’s concrete example is the mcp-antmonitor binding in the Emergency Time-Series Trend Analysis scenario. Around these components, the implementation adds a Context Engineering Engine with Summary-Based Memory Compression, Configurable Context Policies, and Structured Report Distillation. Structured summaries include key findings, confidence scores, and evidence pointers, specifically to reduce information loss and “lost-in-the-middle” effects when a supervisor consumes outputs from many specialists.

4. Evaluation methodology and empirical results

The evaluation is organized around three research questions: the efficiency and effectiveness of multi-agent versus monolithic architectures; adaptability to domain knowledge and different LLMs; and the contribution of architectural components. For RCA, the benchmark combines the public OpenRCA dataset with 335 scenarios and a proprietary AntRCA dataset adjusted to reflect the true distribution of daily SRE tasks. Metrics are a human-graded 100-point scoring rubric and execution time in minutes. The central comparison is between V1, a Basic ReAct Agent; V2, a Phased-Control ReAct Agent; and V3, the Multi-Specialist Agent Framework corresponding to OpenDerisk (Di et al., 15 Oct 2025).

Model	Score (V1 / V2 / V3)	Time in min (V1 / V2 / V3)
Qwen-QWQ-32B	43 / 56 / 74	6 / 19 / 22
Deepseek-R1-0528	45 / 54 / 73	6.5 / 21 / 23.5
Bailing-Deepseek-V3	39 / 58 / 76	2.5 / 6.9 / 9.1

Across all tested models, the score ordering is consistently $V3 > V2 > V1$. The paper interprets this as evidence that multi-specialist decomposition substantially improves diagnostic quality. It also emphasizes the latency trade-off: on Qwen-QWQ-32B, for example, latency increases from 6 minutes in V1 to 22 minutes in V3 because of decomposition overhead, additional reasoning steps, and inter-agent communication.

The adaptability study centers on a trace-based RCA agent evaluated in production over one month across 1,743 developers and 6,000+ real-world cases. The reported aggregate success rate exceeded 80%. Model-wise qualitative results indicate that GPT-oss-120B, GPT-oss-20B, GPT-4o, Deepseek-v3, Qwen3-235B, and Claude-Sonnet-3.5 achieved the paper’s “>80%” threshold with high analysis-quality ratings, while some other models were below threshold or exhibited speed or stability issues. The paper also reports qualitative case comparisons in which V3 provides deeper configuration/runtime analysis, more complete causal chains, better cross-module reasoning, more precise localization, and higher operability than the monolithic baselines.

These experiments do not present OpenDerisk as universally dominant on every metric. Rather, they establish a specific empirical profile: higher diagnostic quality and richer cross-domain reasoning at the cost of additional latency and computational expense.

5. Industrial deployment, extensibility, and open-source status

OpenDerisk is reported as deployed at Ant Group as a general-purpose AI-SRE platform rather than a single-purpose RCA tool. Over a three-month production period, it was applied to 13 new application scenarios; internal developers created over 50 new specialized agents; the system served over 3,000 daily active users; and it executed more than 60,000 diagnostic runs per day. This deployment evidence is central to the paper’s claim that the framework is industrial-grade rather than purely experimental (Di et al., 15 Oct 2025).

The implementation is Python-based and model-agnostic in the pragmatic sense that it primarily used Deepseek, Qwen, Claude, and GPT models. A central Orchestrator manages agent lifecycle, task delegation, and message passing, while inter-agent communication is implemented through an asynchronous message bus to support parallelism and scalability. The framework also includes a web-based user interface and a custom Visualization Protocol that renders the Reasoning Flow, Evidence Chain, and Multi-Agent Dynamics, thereby making investigations auditable at the level of intermediate reasoning and evidence rather than final conclusions alone.

A detailed case study is Emergency Time-Series Trend Analysis. The scenario began with a simple two-agent team comprising Analysis_Agent and Summarizer_Agent, and later evolved into a richer validation setting with DualRun_Orchestrator, isolated TeamMode memory for blind analysis, and Critic_Summarizer_Agent in GroupMode for fair comparison against a legacy system. The case is used to illustrate a key practical property of the framework: developers can assemble new specialist constellations and evaluation workflows without redesigning the core runtime.

The framework is open source at https://github.com/derisk-ai/OpenDerisk/. The paper states that both the framework and the web-based UI are open sourced, and presents this as a basis for reproducibility and further research on diagnostic multi-agent systems.

6. Limitations, future directions, and terminological scope

The paper identifies several limitations. First, OpenDerisk is currently assistive rather than autonomous; final remediation still requires human oversight. Second, diagnostic quality depends heavily on the quality and maintenance of the knowledge base. Third, the most capable multi-agent mode is slower and more computationally expensive than simpler monolithic baselines. Fourth, the reported production evidence comes from Ant Group, which has a mature observability ecosystem, so generalizability to organizations with different stacks or data quality remains unvalidated. Additional failure modes discussed or implied include stale or poor knowledge causing incorrect diagnoses, context-window distortion, JSON parsing problems in some Qwen-series models, and latency in urgent settings (Di et al., 15 Oct 2025).

The roadmap is presented as a transition from a co-pilot to an autonomous “pilot” in a future V4. The stated directions are safe closed-loop remediation; reinforcement learning for individual tool-use optimization and system-level collaboration optimization; a self-healing Knowledge Engine in which agents contribute back to the knowledge base; and an adaptive reasoning controller that dynamically trades off diagnostic depth against latency. These plans are notable because they preserve the framework’s specialist, diagnostic-native character rather than reorienting it toward general autonomous software engineering.

The term “OpenDerisk” also has a broader and less formal semantic perimeter in adjacent literature. It should be distinguished from the unrelated “DeRisk” credit-risk prediction framework (Liang et al., 2023) and from generic optimization de-risking work (Bienstock et al., 27 May 2026). In some later technical syntheses, “OpenDerisk-style” is used as an interpretive label for DeFi risk-reduction agendas, including simulation-based AMM portfolio optimization, oracle-safe parameter synthesis, and derivative-theoretic lending-risk analysis (Nosse et al., 2024, Deng et al., 2024, Szpruch et al., 2024). This suggests a secondary usage in which “OpenDerisk” denotes a broader pattern of structured risk engineering. In the strict encyclopedic sense, however, OpenDerisk denotes the specialized, open-source multi-agent AI-SRE framework introduced for diagnostic reasoning in production systems (Di et al., 15 Oct 2025).