# Coordinate Agents

> Split work across agents — delegate an isolated implementation to OpenAI Codex through codex-bridge in its own git worktree, run independent parts in parallel subagents, then review and integrate the results yourself. Includes the former /codex-frontend workflow. Use for "use Codex for this", "delegate to Codex", "split this work", "run agents in parallel" or "orchestrate agents".

- Skill: `amsborse/coordinate-agents` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx get-skill add amsborse/coordinate-agents`
- Raw SKILL.md: https://api.skillmd.com/api/skills/amsborse/coordinate-agents/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Web & Frontend
- Author: amsborse (https://skillmd.com/u/amsborse)
- Updated: 2026-11-17
- Page: https://skillmd.com/skills/amsborse/coordinate-agents

---


# Coordinate agents

You stay the architect: you write the brief, review every result and integrate it. Delegated
output is untrusted until you have read the diff and run the checks.

## When to split

Only for work that is genuinely independent (separate files or modules, a clear acceptance check)
or a second opinion. Keep tightly coupled work in one agent. Small tasks you can finish in a few
steps stay with you.

## Who gets which work

Follow `~/agent-skills/shared/agent-routing.md`: UI work goes to Codex, simple isolated tasks to
Antigravity (`--agent antigravity`), complex work stays with you. When a run reports
`quotaExhausted`, move to the next agent in that row.

## Delegated implementation (Codex, Antigravity, OpenCode)

Follow `workflows/codex-implement.md` with the agent the routing picked: brief file, its own
worktree, `codex-bridge run --agent <agent> --sandbox workspace-write`, then review and bring the
change back. Bridge rules are in `~/agent-skills/shared/codex-bridge.md`.

## Parallel subagents (Claude)

One goal per agent, a brief that stands alone (files to read, what is ruled out, the output
format), launched in a single message. For edits, give each its own worktree. Record each agent's
task and status in a ledger when there are more than two.

## Integrate

Read each diff, run the tests, resolve conflicts yourself, verify the result (`verify`) and get a
review of large changes (`review-code`). Report which agent did what and anything that failed.

