Descriptions:
Joshua Snyder, a software engineer at PostHog, presented at the AI Engineer conference on a system PostHog is building to eliminate manual dashboard monitoring: an agentic pipeline that detects product anomalies and automatically opens pull requests in GitHub without any human intervention in the loop.
The pipeline begins by ingesting signals from PostHog’s analytics, session replays, error tracking, and logs — handling trillions of events per month. An LLM classifier sits at the top of the pipeline to filter malicious or adversarial inputs from public event sources. Signals are then normalized into a common structure, weighted by importance, and embedded. A key engineering insight Snyder shared is that grouping cross-modal signals (errors, Slack messages, session replays) requires embedding LLM-generated queries about the signals rather than embedding the signals themselves, since different data types cluster separately in embedding space. A research agent built on the Claude Agent SDK runs in a Modal sandbox with access to PostHog’s MCP server, the codebase, and external MCPs like Linear and Notion; it outputs a problem summary, priority score, and Git-blame-derived reviewer list. An actionability step decides whether to open a PR, queue for more evidence, or request human input.
Snyder’s talk is a practical reference for platform and developer-tooling engineers exploring autonomous code-change pipelines, covering specific architectural tradeoffs, safety filtering strategies, and lessons learned from building agentic systems at production scale.
📺 Source: AI Engineer · Published June 10, 2026
🏷️ Format: Hands On Build
