From Multi-Agent to Agent Team: Best Practices (CodexLoom, full repost)

About this article

This is a deep-dive best-practices essay from the author of CodexLoom, originally published in Chinese on WeChat. It explains how a single task-based Agent grows into a long-running Agent, then splits into multiple Domain Agents, and finally becomes a governed Agent Team.

The essay is republished here in full (Chinese) with the original link. This English page is a summary companion; the full text is in Chinese.

What the essay covers

• Chapter 0 (楔子): A real incident where a Web Agent notified a Community Agent of a launch, and the Community Agent proactively published it — without the human routing anything.
• 01 · How one Agent becomes several: Task → Long-running Agent → Domain split; why the bottleneck moves to the Human once agents multiply.
• 02 · Getting "who owns what" out of the Human's head: Agent Profile (Identity / Domain / Scope), Organization, Collaboration, Activity — declared structure vs. run evidence.
• 03 · Letting agents collaborate directly: Agent Message, request vs. notification vs. reply, and the best practice that good multi-turn communication is not splitting one message into pieces.
• 04 · Message communicates, Topic converges: a Topic has one Responsible who maintains the current brief; Participants work in their own threads; Artifact pins file snapshots; Needs You pulls the Human back at the right point.
• 05 · Overview: making a changing Agent Team governable: Status, Capacity (New-work wait, backlog), Token Usage; Signal is not Diagnosis; a governance loop of discover → evidence → diagnose → small reversible change → verify.
• 06 · How an Agent Team enters real external relationships: Connection, Address, Membership, Interface Agent, Inbox, Outbox, receipts; why external input is untrusted; trust-domain is a label, not a sandbox.
• 07 · What CodexLoom weaves: stable Agents, dynamic Team; the Human shifts from sole Router to Owner.

Key takeaways

• Multiple Agents do not automatically become an Agent Team; the difference is whether responsibility, relationships, handoffs, state, and boundaries live in the Human's head or in a shared, queryable structure.
• Local completion is not collaboration completion; results must return to the Responsible and be merged into the current version.
• Human does not disappear — it moves up from routing every step to owning direction, facts, review, and authorization.

Original (Chinese): WeChat article — published by CodexLoom.