Five AI agents walk into an outage

What it does

OpenDeRisk is a multi-agent system that automates root-cause analysis for production incidents. It ingests logs, traces, and metrics—currently centered on Microsoft’s OpenRCA dataset—and coordinates five specialized agents to investigate failures and produce evidence chains. A web UI renders the full diagnostic flow, from alert awareness to final verdict.

The interesting bit

The Code-Agent writes code on the fly to perform the final analysis, turning root-cause investigation from a static report into a live computation. Meanwhile, the Vis-Agent maps the entire multi-agent deliberation and evidence chain in real time, so you can watch the committee think.

Key highlights

• Agent specialization: SRE-Agent, Code-Agent, ReportAgent, Vis-Agent, and Data-Agent each own a distinct phase of the investigation.
• Dynamic forensics: Code-Agent generates analysis code dynamically rather than relying solely on pre-baked heuristics.
• Heavy data appetite: Ships with support for the 20–26 GB OpenRCA dataset and accepts flame graphs, Excel, and custom telemetry.
• Visual evidence chains: The Vis protocol renders the full diagnostic flow and agent hand-offs, not just the conclusion.
• Published framework: Backed by an arXiv paper (2025) with design, implementation, and case studies.

Caveats

• The README notes that only the highlighted components in the architecture diagram are currently implemented, so the codebase appears partial.
• The primary mode requires downloading and decompressing a large external dataset before you can do much.
• Default configuration and proxy settings target Aliyun and DingTalk, which suggests a China-centric happy path.

Verdict

Teams drowning in unactionable alerts and slow post-mortems should evaluate this; if your stack lacks the telemetry volume or you need a fully baked end-to-end product, it may still be too early.