Agent Teams Creator

Turn multi-agent swarm ideas into an explicit team runtime with task state, mailbox protocols, and verification boundaries.

What It Is

agent-teams-creator is a Codex skill for analyzing, designing, and implementing structured agent team runtimes.

It helps with requests like:

• "How should an agent team runtime actually work?"
• "Design a coordinator + workers system for this project."
• "Add a shared task board and structured inter-agent messaging."
• "Make our multi-agent runtime safer and easier to reason about."

Why It Matters

Most agent-team designs go wrong because they confuse three different things:

• shared state
• transport
• role authority

This skill forces those apart.

Its central model is:

• task board = state
• mailbox = transport
• team lead / coordinator = control plane

That distinction makes the runtime much easier to build, test, and debug.

What You Get

• A clean model of team bootstrap, spawn modes, and teammate identity.
• A protocol view of inter-agent messaging.
• A task-board-first coordination model.
• Clear guidance on default isolation vs optional worktree isolation.
• Strong verification rules so implementation workers do not self-certify success.

Key Capabilities

• Team file and deterministic identity design.
• Shared task board architecture.
• Structured mailbox message catalog.
• Spawn backend model: in-process, pane-based, and fallback execution.
• Coordinator/lead behavior model.
• Isolation and cleanup guidance.
• Independent verifier path.

Validated Value

This skill was pressure-tested in practical Codex workflow scenarios.

Observed difference:

Scenario	Without the skill	With the skill
Explain and rebuild a protocol-driven agent team runtime	The agent often stopped at "I inspected these files" or broad swarm commentary.	The agent consistently described the coordination spine, protocol message types, task-board model, spawn backends, isolation stance, and verification rules.

That means the skill materially improves answer shape, not just wording.

Before / After

Without agent-teams-creator

• answers can stall at file-reading summaries
• "multi-agent" gets described too loosely
• the difference between task board, mailbox, and coordinator is often blurred
• worktree isolation is easy to overstate or misstate

With agent-teams-creator

• the answer is forced toward the real team runtime
• the task board is identified as shared state
• the mailbox is identified as transport and protocol bus
• the lead is identified as control-plane authority
• default shared-workspace collaboration is separated from optional worktree isolation
• verification is treated as an independent phase, not worker self-certification

Case Study

Case: "Explain an agent team runtime and how to build one"

Baseline behavior

• one baseline run stopped at "I inspected these files"
• another broad answer still risked collapsing protocol details into generic swarm language

With the skill

• the answer explicitly covered:
  • coordination spine
  • structured message types
  • spawn modes
  • teammate identity
  • task-board ownership/blockers
  • coordinator role
  • isolation stance
  • verification stance
  • facts vs inferred behavior

Why that matters

That moves the output from "interesting analysis" to "something you can actually build from."

Quick Start

Clone or copy this repository into your Codex skills directory:

Windows

git clone https://github.com/Arthurescc/agent-teams-creator `
  "$env:USERPROFILE\\.codex\\skills\\agent-teams-creator"

macOS / Linux

git clone https://github.com/Arthurescc/agent-teams-creator \
  "${HOME}/.codex/skills/agent-teams-creator"

If you use CODEX_HOME, install it under:

$CODEX_HOME/skills/agent-teams-creator

Restart Codex after installation so the skill list refreshes.

Example Prompts

• Use $agent-teams-creator to explain how a protocol-driven agent team runtime should work.
• Use $agent-teams-creator to design a coordinator-plus-workers system for this repo.
• Use $agent-teams-creator to add a task board and mailbox protocol to our multi-agent runtime.
• Use $agent-teams-creator to compare our current swarm design against a stronger team-runtime model.

What It Covers

• team bootstrap
• teammate identity
• spawn modes
• mailbox protocol
• task board model
• coordinator role
• isolation stance
• verifier flow
• facts vs inferred/gated behavior

Repository Layout

SKILL.md
agents/openai.yaml
references/
  agent-teams-mechanism.md
  agent-teams-workflow.md
assets/
  icon-small.svg
  icon-large.svg

Development

This project is independently designed and developed for practical Codex skill workflows.

It is focused on making multi-agent runtime design more explicit, protocol-driven, and implementation-ready.

Who This Is For

• agent-platform engineers
• developer tooling teams
• researchers building multi-agent coding systems
• teams adding verifier or reviewer workers
• anyone who wants a protocol-driven alternative to vague swarm prompting

Chinese Documentation

See README.zh-CN.md.

More Skills

Browse the collection page: codex-skills-hub

License

MIT