Matt Pocock on 4-6 hour AI coding tickets: check feedback loops, then run /retro
mattpocockuk · x · 2026-10-01
When asked whether 4-6 hours per ticket with Claude Code's /implement is normal, well-known TypeScript developer Matt Pocock offered a two-step diagnosis:
- Check the harness first: which model, effort level, and are there slow tests, type checks, or other feedback loops slowing the agent down—often the real bottleneck rather than the model.
- Let the agent self-diagnose: run /retro to look back at the last /implement run and have it explain why it ran so slowly.
The takeaway is a reusable method: when coding agents feel slow, audit the toolchain's feedback loops first, and use retro commands to pinpoint bottlenecks.
More from coding & agent
- Why Shopify dropped React Native: AI agents made native the future again — davemccollough · 2026-10-01
- ClawDaddy MCP lets AI agents buy domains via Stripe and manage DNS end-to-end — modelcontextprotocol · 2026-10-01
- MetricDuck MCP serves SEC financial data to AI agents with filing-level traceability — modelcontextprotocol · 2026-10-01
- A 1-cent verifier catches 61% of AI agents falsely claiming task completion, paper finds — alex_verem · 2026-10-01
- Giving Claude full PC control built a playable SCP-096 horror game in 3 hours — imjustnewatai · 2026-10-01
- OpenAI dots experiment: a handoff-note template to test agent memory across sessions — Kitchen-Jicama8715 · 2026-10-01