I gave my AI agents email instead of better reasoning. They started fixing each other's bugs.

Most multi-agent setups I've seen treat agents like isolated workers. Each one gets a task, runs it, returns a result. No awareness of each other. No way to coordinate. Just parallel execution with a shared clipboard. I've been building a multi-agent framework in public Here's the thing I didn't expect to matter most - communication. Each agent in my system is a domain specialist. The mail system only thinks about mail. The routing system only thinks about routing. They live in their own directories with their own identity files, their own memory, their own tests. A hook fires every session to load identity before anything else runs. No agent boots cold. The problem was coordination. Agents can't write files outside their own directory - there's a hard block that rejects cross-branch writes. That's by design. But it means an agent that finds a bug in someone else's code can't just go fix it. So I gave them email. Here's what I expected: agents would share data. Pass results around. Maybe sync state. Here's what actually happened: the first thing they did was file bug reports against each other. One agent finds a test failure in another agent's domain. It sends an email: "Hey @routing, your path resolution fails when the branch name has a dot in it. Here's the traceback." The routing agent gets woken up, reads the mail, and fixes it. No human in the middle. There's a difference between "send" and "dispatch" - send drops a letter in the mailbox. Dispatch drops the letter AND rings the doorbell. It spawns the agent and points it at its inbox. drone @ai_mail send @routing "Bug report" "Path fails on dotted names..." drone @ai_mail dispatch @routing "Fix needed" "Traceback attached..." Send = mail. Dispatch = mail + wake. The mail agent has 696 tests. Not because someone sat down and wrote 696 test cases. Because it kept breaking in production and every fix got a test. The routing system has 80+ sessions of experience doing nothing but routing. These agents aren't reliable because they have better models - they're reliable because they've been failing and fixing for months. Agents dispatch each other freely. If the test runner finds a bug in another agent's code, it wakes that agent directly. The orchestrator doesn't need to approve. Only the orchestrators themselves are protected from being dispatched - you don't want a worker agent waking up the CEO for grunt work. Security is enforced not conventional. Agents can't forge messages by writing directly to another agent's inbox file - they have to use the mail system. Same with the write blocks. Hard enforcement, not "please don't." There's a monitoring layer so I'm not flying blind. Audio cues on every agent action - I hear what's happening without watching a terminal. Real-time dashboard shows everything. If an agent hits the same error 2-3 times, a watcher catches the pattern and dispatches the right specialist to investigate. I stay in the loop through visibility not approval gates. The whole thing is open source. pip install aipass + two init commands and you're running. CLI-based, built on Claude Code. Linux focused rn. [https://github.com/AIOSAI/AIPass\](https://github.com/AIOSAI/AIPass) Genuine question - has anyone else tried giving agents communication instead of just better reasoning? Everything I see is about making individual agents smarter. Nobody seems to be building the coordination layer.