Why we stopped using an automated SRE agent
9 hours ago
- #AI pitfalls
- #SRE automation
- #context-driven AI
- The team stopped using an automated SRE agent due to issues with confident hallucinations and worsened outages.
- Three fatal flaws were identified: signal dilution from information overload, reliance on outdated documentation via RAG, and the 'confused deputy' risk from granting write access.
- The solution shifted to a lightweight, local-first approach using tools like NeatContext, which allows explicit bundling of relevant data for LLMs.
- This new paradigm enforces strict system boundaries, injects tribal knowledge, and ensures safety by keeping engineers in the loop.
- The change resulted in reduced MTTR, elimination of hallucinations, and a trusted assistant for on-call engineers.