a builder's codex
codex · operators · Claire Vo · ins_agents-as-team-not-tools

Agents work when treated as a team, not a single super-tool

By Claire Vo · 3x CPO; founder of ChatPRD; host of How I AI · 2026-04-28 · podcast · Claire Vo on running 9 AI agents — Lenny's Podcast

Tier A · TL;DR
Agents work when treated as a team, not a single super-tool

Claim

Don't throw every task at one AI agent. Build one agent per role, each with its own context window, identity, and tool scope, and manage them like teammates.

Mechanism

Context is the bottleneck for agent quality, not raw model capability. A single agent asked to do nine jobs accumulates conflicting context across them and degrades on each. Splitting work across role-scoped agents keeps each agent's context window clean, lets you onboard each to one job, and lets you grant tool access per-role rather than per-account.

Conditions

Holds when:

Fails when:

Evidence

"Where people stumble with OpenClaw is they think they can throw any task at a single agent and get great results."

"I have nine Slack channels. My marketing team's in one, sales in another, dev in another. My development team does not care what was posted on X today."

Claire ran nine agents on three Mac Minis: Polly (work EA), Finn (family), Sam (sales SDR), Howie (podcast research), Sage (course PM), Q (kids' tutor), and others. Sam-the-SDR replaced 10 hours/week of contractor work running PLG sweeps and drafting soft outreach.

— Claire Vo on Lenny's Podcast, 2026-04-28

Signals

Counter-evidence

A single very strong general agent (e.g., long-context Claude with broad tool access) can outperform a poorly-managed team of small agents. The win is in the management discipline, not the architecture itself. Operators who can't maintain identity + tool docs may get worse outcomes from the team approach.

Cross-references

Open the interactive view → View original source → Markdown source →