Original Summary

Most multi-agent setups treat agents as isolated workers. Each one gets a task, runs it and returns a result, with no awareness of the others and no way to coordinate. AIPass is an open source framework built in public for about seven months: 18 agents, over 21,000 tests, 280+ stars on GitHub. The part that turned out to matter most was not reasoning. It was communication. Specialists in their own directories Each agent is a specialist in one domain. The mail agent only thinks about mail. The routing agent only thinks about routing. Each lives in its own directory with its own identity file, its own memory and its own tests. A hook loads that identity at the start of every session, before anything else runs, so no agent starts cold. That created a coordination problem. An agent cannot write files outside its own directory. It is a hard block, not a convention: the write is refused. So an agent that finds a bug in someone else's code cannot just go and fix it. So the agents got email The expectation was that agents would use mail to share data, pass results around, maybe sync state. What the developer saw instead was that the first thing they did was file bug reports against each other. One agent hits a failure in another agent's domain and writes something like: "your path resolution fails when the branch name has a dot in it, here is the traceback." The owner reads it and fixes it, with no human in the middle. There are two ways to send. email drops a letter in the mailbox. dispatch drops the letter and rings the doorbell: it starts the receiving agent and points it at its inbox. drone @ai_mail email @drone "Bug report" "Path fails on dotted names..." drone @ai_mail dispatch @drone "Fix needed" "Traceback attached..." Email is mail. Dispatch is mail plus a wake. Reliability came from breaking The mail agent has about 1,460 tests. Nobody sat down and planned 1,460 test cases: it kept breaking, and every fix got a test. The routing agent has more than 110 recorded sessions of doing nothing but routing. These agents are not reliable because they run a better model. They are reliable because they have been failing and fixing for months. Who can wake whom Agents dispatch each other freely, and nobody has to approve it. The exception is the managers: a dispatch from another agent to a manager is delivered as mail and does not wake it. A worker should not be able to wake the orchestrator for grunt work. The rules are enforced, not requested. An agent cannot forge a message by writing straight into another agent's inbox file: that write is blocked, and the only way in is the mail system. The same goes for the directory blocks. Seeing it run A monitoring agent logs every branch in real time, and hooks can play a sound on agent actions, so you can hear work happening without watching a terminal. A watcher read

  • 情报分类：商业与市场研究
  • 分类依据：内容涉及商业、投资或市场动态
  • 信息来源：Reddit · SideProject
  • 发布时间：2026/10/4 10:19:43