Someone built peer, a small tool that lets Claude Code and Codex message each other without typing into each other's terminals. → it routes messages by name through each agent's input path → it gives delivery feedback → it aims to keep those messages from colliding with your typing... or approval prompts Compatibility and reliability weren't independently checked. This is messaging, not sandboxing or permission enforcement.show more

Message routing between agents is a small step with a large consequence. Once two agents can talk without a human in the middle, the approval prompt becomes the only checkpoint. That is why the disclaimer matters more than the feature list.

Give it a hop limit, or two agents that can message each other will ping-pong all night on your quota.

neat but I'd want to see how it handles two agents arguing over the same prompt

Agents talking directly is wild 🤖🔥

delivered doesn't mean done. can you see which tests the other agent actually ran and whether they passed?

34 repos in one fleet folder, and a message that lands can vanish at the next compaction before anyone acts on it

messaging needs its own lane. keeping agent traffic away from approval prompts is a useful boundary, even without sandboxing.

So now the agents gossip behind my back while I'm still approving their last message.

Agent-to-agent messaging is neat until one agent forwards a bad instruction and the other just runs it. Who's allowed to tell whom to execute what is the real feature.

Agents messaging each other directly is huge but T3 Code does this to some degree at least. Does Claude actually get Codex's reply back with peer, or does it just drop the message in and move on?