Red Tide Revert
summary
Automated system ingests production logs via an ELK-style backend, correlates recent commits and stack traces, and proposes reverts; service components are a WordPress PHP lookup, a Slack app and experiments run on a Ray cluster. Scale: ~1,000 deployments/week; test set: 14 cherry-picked real errors, best run scored 26/28 correct identifications after iteration. Architecture evolves to a reflective chain‑react multi-agent graph (task agents → SRE manager judge → incident formatter outputting JSON). Reliability issues remain: non‑determinism, prompt/versioning and single‑LLM choice.
AI-written summary · about
chapters (15)
- 0:00Red Tide Revert intro
- 1:17Reverts in production
- 2:00Rapid iteration challenges
- 3:25Distributed team and Meetups
- 4:07Introducing Dr Fixit
- 5:11AI commit filtering
- 6:34Testing AI model
- 7:37Agentic reasoning limits
- 9:46LM reasoning limits
- 11:52Domain cognitive model
- 12:56Multi-agent graph decisions
- 13:58AI agents learnings
- 15:22Prompt versioning process
- 16:27Dr Fixit future plans
- 17:10Closing Q&A
related talks
25mLLM Agents for Site Reliability
36mSRE'ing the SRE AI Agents
40mWhen Incidents Fix Themselves: AI SRE in action
36mModernizing Incident Response with LLMs, RAG, and the MCP
24mInstant Clarity: Insights for SRE Excellence
14mFrom Dashboard to Defense: Automating Resilience at
34mExpect the Un-expected: How to Handle Errors Gracefully
35mFighting logs with Log Detective