Self Driving Products: Product Signals to Pull Requests
summary
A pipeline ingests PostHog signals and claims to open working PRs so engineers wake to green merges not dashboards. It uses an LLM classifier for safety, normalizes events, and clusters by embeddings of LLM-generated queries instead of raw signals. A research agent running Claude in Modal pulls logs via an MCP server, uses repo context and Git blame, decides actionability (errors are usually actionable; Slack and replays rarely), and loops CI until green while learning that turning repeated agent runs into one-shot calls collapses cost.
AI-written summary · about
chapters (10)
- 0:00Self-driving products intro
- 1:59Slow observability workflow
- 2:41Vision for auto PRs
- 3:44Pipeline overview
- 5:52Challenges in grouping signals
- 7:17Research agent tools
- 9:01Actionability of reports
- 10:06Fixes and CI integration
- 11:07Pipeline lessons learned
- 13:53Current status and goals
related talks
36mAgentic CI/CD: from Pull Request to Production without the Paper Cuts
23mFrom Blind Spots to Merged PRs: Continuous Agentic Performance Optimization
21mFrom Signal to PR: Anatomy of a Self-Improving Agent
31mBeyond the Copilot: A Blueprint for Autonomous Mobile Reliability
27mIntegrating Agentic AI Into Internal Developer Platforms To Minimize Errors and Downtime
38mEthical Artificial Intelligence to Advance Welcoming Communities and Identify Potential Communica...
20mThe Missing Layer After Launch
12mHow to maintain a codebase – when everyone can code