The bridge across agents — delegate tasks to another AI agent running headless.

agent-bridge is an agent skill that lets an orchestrator agent delegate tasks to another AI agent running headlessly. It assembles a structured prompt, optionally shows it for review, then spawns the target CLI and returns the result.

Supports two modes:

• Review mode — compile → inspect → execute
• Auto mode — compile + execute in one step, no review

Via npx skills:

npx skills add davipedro/agent-bridge

Via Claude Code plugin marketplace:

/plugin marketplace add davipedro/agent-bridge
/plugin install agent-bridge@agent-bridge

Or copy skills/agent-bridge/ into your project's .claude/skills/ directory.

• Node.js 18+
• claude and/or gemini CLI in PATH and authenticated

Invoke with /agent-bridge in Claude Code. The skill guides you through:

1. Harness (claude or gemini), model, execution type, working directory, task prompt
2. Prompt review (or skip with auto mode)
3. Execution and result

{
  "action": "run",
  "harness": "gemini",
  "model": "gemini-2.5-pro",
  "executionType": "analysis",
  "sessionId": null,
  "cwd": "/absolute/path/to/project",
  "prompt": "Find why auth.py fails on expired sessions"
}

{
  "action": "compile",
  "harness": "gemini",
  "model": "gemini-2.5-pro",
  "executionType": "analysis",
  "sessionId": null,
  "cwd": "/absolute/path/to/project",
  "prompt": "Find why auth.py fails on expired sessions",
  "entryPoints": ["src/auth.py"]
}

Then pass promptCompiled to action: "execute" after reviewing.

Pass sessionId from a previous execute/run output to resume the same agent session. Set to null to start fresh.

The bridge script is at skills/agent-bridge/scripts/bridge.ts. Run directly with:

echo '<input-json>' | npx tsx skills/agent-bridge/scripts/bridge.ts