Joshua Snyder is a PostHog product engineer and TimeNavi co-founder building systems that turn product analytics, customer feedback, and software errors into working code. His approach to self-driving product development replaces passive observability dashboards with agents that investigate real user problems and open reviewable pull requests.
Snyder studied at the University of Cambridge, researching embeddings for dynamic network graphs and building a reinforcement-learning crawler for academic websites. His software projects include DeepWalk and SpeedStore, a Google Apps Script storage library that organizes properties into chunks and supports compression.
After graduating, he worked full time on TimeNavi, the calendar-based time-management company he co-founded. His argument for modular monoliths reflects firsthand startup priorities: define service boundaries only when the product is mature enough to justify separate deployment, monitoring, and operational overhead.
In 2023, he contributed to GPTeam, an open-source multi-agent simulation whose characters use memories, plans, directives, and reflection to decide how to act. At PostHog, he initially worked on growth, where reviewing session recordings of customers struggling with setup helped inspire the company’s AI onboarding wizard. PostHog reported that wizard users converted to paid usage at 14.2%, compared with 2.6% for other users, and reached their first tracked event in approximately half the time.
How he designs autonomous product engineering
- Convert product signals into actionable code. Snyder’s automated improvement pipeline, described with colleague Cleo Lant, combines analytics, errors, session replays, experiments, and customer feedback. Research agents investigate significant problems before coding agents propose GitHub pull requests; insufficient evidence triggers further investigation, while ambiguous product decisions return to humans.
- Group problems by meaning, not format. Ordinary embeddings can cluster stack traces with other stack traces and customer messages with other messages, even when both describe the same broken checkout. Snyder uses LLM-generated semantic queries to connect structurally different evidence describing a common product failure.
- Constrain autonomy with security and context. Customer-controlled signals undergo prompt injection screening before agents investigate inside Modal sandboxes using the Claude Agent SDK. MCP connections provide observability data, repository context, and information from systems including Linear and Notion; snapshotting lets coding agents resume after failed checks or review comments.
- Evaluate realistic outcomes before optimizing costs. Snyder insists on production-representative agent evaluation, clear actionability thresholds, and verification that proposed changes address actual customer problems. He experiments with capable agents first, then simplifies repeated behaviors into cheaper model calls once their useful patterns become clear.
His longer-term goal is a closed-loop software improvement system that runs experiments, ships appropriate fixes behind feature flags, measures results, and learns from accepted changes, rejected suggestions, deployment failures, and resolved production errors.
