PostHog engineer Josh Snyder describes a pipeline that converts product observability signals into automated pull requests. The system ingests trillions of events monthly, normalizes and groups signals into reports using LLM-generated query embeddings (rather than raw embedding similarity), runs a Claude-powered research agent with MCP tool access, assesses actionability, and finally executes code fixes in a sandboxed environment. Key lessons include: evals on representative production data are essential, embedding models cluster by structural similarity not semantics so queries should be embedded instead, agents need specific problem descriptions to avoid noisy PRs, and cost optimization should come after experimentation by observing agent behavior patterns. The alpha feature aims to let developers wake up to green PRs rather than dashboards full of errors.