Let me be honest about how this started: I read about people running multiple AI agents at once and thought, I want that. I’m not an engineer, but I had several machines — laptops and rented cloud servers alike — and stubbornness, so I tried to build a small fleet — a handful of agents that could research, write, review, and code in parallel.
The first thing I learned is that the obvious worry isn’t the real one. I assumed the AI providers would rate-limit me — that running many agents at once would hit a wall. They mostly don’t, at my scale. The actual problem is subtler: one specific AI model handles one request at a time well. If two of my agents call the same model at the same moment, they degrade each other — both get slow or flaky.
So the core of what I built isn’t glamorous. It’s a reservation system. Before an agent starts, it claims a specific model, and the dispatcher won’t let another agent take that same one until it’s free. If the model someone wants is busy, the system suggests a different agent on a different model that can do the same job. Think of it like a kitchen with one of each pan — you can’t have two cooks using the same pan at once, so you assign pans, not just cooks.
Three walls I hit, in case you’re attempting anything like this:
Parallel work needs separate workspaces. I had multiple agents writing to the same folder of files at once and they collided — one overwrote another’s work. The fix was a git feature called worktrees that gives each agent its own workspace sharing one underlying project. Cheap, instant, no more collisions. I didn’t know worktrees existed until this broke.
Don’t mix systems that work differently. I tried to have my research fleet and my coding tools share the same setup, and it kept breaking — they authenticate differently, store settings differently, fail differently. Keeping them deliberately separate fixed more than any clever integration did.
Every tool stores its settings in a different shape. Five agent tools, five completely different config file formats for the same list of AI models. I ended up building one master list plus a small translator for each tool, with backups before every sync — because I learned the hard way that a bad sync can brick the whole fleet at once.
The unglamorous truth about running multiple agents: the exciting part is the AI. The part that breaks is the plumbing. I spent far more time on the plumbing.