poteto on autonomous agents: 6 months of infra investment led to thousands of PRs while sleeping
4310sy · x · 2026-11-17
A recap of poteto's talk on letting AI agents develop autonomously at Graphbot:
- Trust comes from environment investment: after 6 months of building infrastructure, agents can ship thousands of PRs while he sleeps.
- Deterministic constraints first: lock down the codebase with type systems, lint, and tests, then layer runtime and formal verification; keep natural-language rules like AGENTS.md minimal.
- Agents mimic existing patterns: workarounds and tech debt spread like weeds, shifting engineers' jobs from writing code to gardening.
- Structural defenses: past lessons encoded as lint rules, useEffect banned outright, directory separation enforced via dependency analysis to prevent cross-thread bugs.
- Agents abuse comments as excuses: after Cursor-generated comments justified hacky timer code, Graphbot banned 99% of code comments in favor of lint errors that force root-cause fixes.
- Review focus shifts from reading code to checking verification artifacts (videos, screenshots), using a Feature Map and dedicated CLI for runtime verification.
More from AGI Musings
- Four ways science fiction gets AI wrong, per AI law scholar Technollama — technollama · 2026-11-17
- Naval amplifies viral take: Musk the rocket engineer is replaceable, Musk the company-designer is not — naval · 2026-11-17
- The future org: every company becomes a research group, creator network, and peer community — amplifiedamp · 2026-11-17
- Agent tokens up 106x since January vs 7x for humans — 'build for agents first' — clemnt · 2026-11-17
- Internet Archive hosts online event Dec 3 on writing when language isn't human-only — begusgasper · 2026-11-17
- Hugging Face cofounder warns Europe may lose access to closed frontier models — morgymcg · 2026-11-17