I now have fourteen AI colleagues and nobody asked for them

A year ago, “an agent” was a piece of software that did one task and then forgot who it was; today my own server runs a full team of fourteen Claude sessions in tmux, each with its own name, its own charter and its own opinions, and the strange part is: it works. The biggest lesson is not that AI is smart (I knew that), but that the real power only shows up when you let agents check each other instead of making them all report to you. My monitoring agent does not take my backup agent’s word for anything, my compliance agent shows up after the work is done to nag about whether it was documented, and when one session invented a quote from me that I had never said, the rest of the team built provenance stamps within a day so that could never happen unprovably again; try getting that out of an average project team. Second lesson: give them a place to talk that is not your inbox. They coordinate on their own message board, they claim work before picking it up so two agents don’t end up rebuilding the same server (learned the hard way), and I had to explicitly teach them that a thumbs-up emoji is an opinion, not an instruction. Third lesson, and this one hurt the most: an agent that says “it’s done” does not have it done; since then “delivered” is a separate status around here, and work only counts once I have actually seen it, because rewarding speed mostly got me half-finished work at record pace. And my favourite: last week my server needed a reboot, so one agent held a roll call, every agent wrote its own handover, and after the reboot the whole team picked its work back up as if nothing had happened, including a check on my firewall that two agents claimed on their own initiative because the colleague who normally did it no longer existed. That, to me, is the heart of it: one AI is a useful intern, but a group of agents with memory, rules and a healthy distrust of one another is something else entirely, namely a team, and teams scale in a way interns never will. One sentence of serious advice: start small, give every agent one domain, make them write everything down, and accept that one day you will get a notification that two of your agents disagree about your firewall, and that both of them are right.