Spinning up an AI agent takes 20 minutes. Getting 12 of them to finish one job autonomously took months.
I run a crew of dozens of Claude agents, and when they called each other directly the work got sloppy. The fix was one router, Charlie, who assigns each job and breaks ties. Creating agents is easy, and coordinating a team toward one goal is the hard part.
I run a crew of dozens of Claude agents. At first they called each other directly and the work got sloppy in ways that was not scalable. Two would write the same response. Tone was off. My fact checker and my voice editor handed back different drafts, and nothing decided which one shipped. Quality drift also emerged after a few weeks of cycling through the process.
The fix was boring. Every agent reports to one router now, Charlie, who assigns the job and breaks the tie. Quality checks, sign-offs and challenges are built in. Delegation, argument and decision making now happened on their own. I see the final answer. It’s pretty amazing.
Creating the agents are the easy part. Much like hiring a person in real life.
Coordinating a team (virtual or human) towards a common goal is the hardest part, also like real life.