OpenClaw, Claude Code, and the Future of Software | Peter Yang on The a16z Show
Peter Yang (Roblox PM, creator) argues coding agents will eat all knowledge work and future companies will stay radically small by design.
- Yang runs OpenClaw daily via Telegram voice, but says its default single-file memory system forgets too often — he replaced it with a 3-layer semantic search setup.
- He prefers Claude Code for synchronous flow-state work, Codex for serious accuracy-critical builds; both feel like slot machines due to variable output and timing.
- An unnamed AI-native vibe-coding company is actively rebuilding its own SaaS stack internally to eliminate subscription payments.
- Task-completion apps (Calendly-style) are most threatened by agents; entertainment and social apps are safer because users open them to feel something, not complete a task.
- Full 100% job-function automation (e.g. customer support via Sierra/Happyrobot) remains rare — most AI products provide large lift but still need humans for the last 10%.
- Yang’s prediction: future product teams shrink from 10 people to 2-3 people plus agents, and founders will deliberately stay small to avoid alignment overhead.
- Peter Steinberger, OpenClaw’s creator, is now at OpenAI and will likely productize personal-agent features inside ChatGPT.
- Yang frames the IDE shift as moving from a making tool to a thinking tool — build naively, hammer the agent, then ask it what it would have done differently and restart.
2026-04-06 · Watch on YouTube