Managing multiple coding agents: hourly-granularity tasks and e2e-only acceptance
dotey · x · 2026-10-28
Stephan shares his playbook for running multiple coding agents in parallel, with宝玉 adding that everyone will become a 'manager of agents':
- Slice tasks at hourly granularity: micromanaging agents at minute-level makes you the bottleneck. Invest in design-phase discussion, document a consensus spec, then let agents run while you brief the next one.
- Acceptance must be real: dozens of green unit tests prove nothing. Use pstack's meta-skills /create-verification-skill (builds e2e test infra + a feature map for real e2e testing) and /create-formal-verification-skill (formal verification for un-enumerable cases and concurrency races), plus meta-skills to keep these updated as the project evolves.
- Communicate asynchronously: don't context-switch the moment an agent finishes; batch like CPU time-slicing.
宝玉 argues multi-loop is inevitable—human attention is scarce while agent execution is async—and agents can now self-correct given clear acceptance criteria and tools, even using multimodal frame analysis to verify video edits.
Related event: Engineers Debate the Shift to Managing AI Agents as the New Manager Role(4 posts)→
More from coding & agent
- Will DePue pitches OpenAI: add API-style fan-out so multi-agent systems dispatch 1000s of tiny cheap models — willdepue · 2026-10-28
- Early users praise DeckWerk after weeks of using it for real talks — vincesitzmann · 2026-10-28
- Agentic video editing goes pro: senior editors 5x output, costs to collapse — dotey · 2026-10-28
- Software is going agent-first: sites, browsers, even a new Rust-like language built for agents — mobileraj · 2026-10-28
- HABIT-Bench: Best agent memory system scores just 33.61% at inferring user habits — SJTU · 2026-10-28
- Regehr's 'Piano Test': would you trust formally verified code with your life? — moyix · 2026-10-28