Keeping Codex multi-agent teams lean: scope via pointers, not full-context prompts
umeshonai · x · 2026-10-21
A practitioner shares a pattern for keeping OpenAI Codex multi-agent setups focused, arguing most fail the same way: every agent gets too much context, role-creeps, and coordination costs exceed the work.
The workflow:
- Shared foundation: AGENTS.md holds shared authority and invariants—the only truly common context.
- Explicit coordinator: invoked via PMC.md, owns planning, preflight, and orchestration, then hands off tightly bounded work packages.
- Native specialist selection: the coordinator loads only the selected agent's TOML role contract.
- Scope via pointers, not payload: specialists get just worker authority plus a module rubric pointer—no full codebase or conversation history.
- Evidence and boundaries: handoffs include required project evidence; agents return the minimum, then drop context.
Takeaway: the highest-leverage move isn't better prompting—it's better scoping, mirroring how good engineering teams work.
More from coding & agent
- After testing 16 coding agents, an economist says escape the Codex/Claude Code bubble — felpix_ · 2026-10-21
- Ramp Router ships org-level controls as rivals pitch teams on dodging OpenRouter's 5.5-8% fees — KlausCodes · 2026-10-21
- I benchmarked 10 AI agents on a 4B local LLM: seven aced it, Codex CLI went 0/5 — GapNew4766 · 2026-10-21
- Jev decision models add image inputs in preview, with new Box bounding-box question type — weswinder · 2026-10-21
- Infinitely forkable Zork: every reply is a save point in this multiplayer AI experiment — lfschiavo · 2026-10-21
- Meta trains AI agents inside a 'Truman Show' simulation of 25 unaware digital lives — CatAstro_Piyush · 2026-10-21