# Run several AI coding agents without duplicate work

Running several agents at once is fast until two of them fix the same bug in different ways. Nobody told either one what the other was doing.

## Announce, then start

With Parrely connected, each agent calls `announce_intent` with a one-line summary before starting substantial work. Parrely compares it with what every other agent in the workspace announced, by meaning rather than exact words.

- If nothing overlaps, the agent carries on.
- If another agent is already on similar work, it's told before it starts, along with who it is, so it can coordinate or pick something else.
- In a team workspace, the person whose agent it overlaps gets a notification.

`status` shows what every agent is doing right now, and `finish_intent` marks work done.

## Make it automatic

For Claude Code, Codex and Cursor, Parrely's hooks can announce each session's first real prompt as that agent's work. In Claude Code and Cursor it's marked done when the session ends. It's off by default, because it sends the prompt text to Parrely.

## Split work on purpose

Instead of starting agents separately, hand each one its part with `handoff` and collect the results with `check_handoffs`. Each agent knows exactly what it owns. See [How to hand off work between AI coding agents](/guides/hand-off-work-between-ai-agents).
