Root Cause Analysis Agent
A parallel agent swarm that investigates equipment failures across an institutional knowledge graph, the public web, and internal documentation, with human-in-the-loop plan review.
Why it matters
When a piece of equipment fails, the answer rarely lives in one place. There’s institutional knowledge captured in a structured failure-mode database, industry standards on the public web, and internal documentation tucked inside datasheets and drawings. The RCA agent runs all three searches in parallel, pauses to let the user inspect and edit the analysis plan before dispatch, and synthesizes the results into a coherent failure story.
Capabilities
- Agent-swarm pattern — several subagents per source type run in parallel and fan their results back into a shared result store.
- Multiple graph-search strategies over the failure-mode database: tag-based queries, semantic vector retrieval, hybrid, and LLM-generated structured queries — chosen per question.
- Human plan review via a first-class interrupt; users approve, edit, or reject the plan before any source is queried.
- Configurable source mix — any one source type can be dialed down to zero to skip it entirely, distinguished from “not set” via explicit sentinel checks.
- Self-reflective subagents — query → search → reflect loop, with reflection depth controlled by a three-tier effort setting.
What makes it hold up
The trap with swarm architectures is hallucinated confidence — ten subagents all agreeing because they all saw the same weak signal. The synthesis step is specifically designed to surface disagreement and flag which findings are supported by which source, so the operator stays in the loop on trust, not just in the loop on approval.
Enterprise project. Official writeup and demo link will be added once online.