I’m curious how other people are finding multiple agents, because I’ve gone all the way around the loop and come back to one.
At one point I had as many as 180 individual agents in my OpenClaw. Then I experimented with 10. Then 8, then 5, then 4. Now I’m all the way down to a single “Agent” and one “Watcher” — which is an agent too, I suppose.
Here’s how it works. The Watcher does one thing. It notices when the Agent has finished a task, and it flips the Agent’s state between Planner and Worker. Planner makes the plan. Worker does the plan. The Watcher merely toggles. It doesn’t think, it doesn’t judge, and it doesn’t coordinate anything. It’s a light switch, not a manager.
I put “Agent” in quotes because it’s really an agent capable of spawning ephemeral sub-agents on whatever LLM it chooses, or on several at once. So the capacity for a crowd is still there. It’s just summoned for a job and dismissed when the job is done, rather than standing around as 180 permanent personalities waiting for something to do.
For anyone who hasn’t followed the OpenClaw story: it’s the open-source assistant that runs on your own hardware and connects your chat apps to AI coding agents, and its docs describe it as “built for coding agents with tool use, sessions, memory, and multi-agent routing.” Peter Steinberger first published it in November. It went through the names Clawdbot and Moltbot in January, after a legal challenge from Anthropic over the first one — and by early March it had passed 247,000 stars on GitHub. Multi-agent routing is a first-class feature. That is exactly why 180 was so easy to reach.
I’m not the first to end up here. Anthropic’s guide to building agents, back in December 2024, put it plainly: “we recommend finding the simplest solution possible, and only increasing complexity when needed,” and it warned that agentic systems “often trade latency and cost for better task performance.” Last June, Cognition published a piece bluntly titled “Don’t Build Multi-Agents”, built on a line I keep returning to: “Actions carry implicit decisions, and conflicting decisions carry bad results.” Their answer was a single-threaded agent with continuous context. The very next day Anthropic described its own multi-agent research system and was candid about the bill: agents use about four times the tokens of a chat, and multi-agent systems about fifteen times. Their conclusion was that the task has to be valuable enough to pay for that.
What I’ve landed on looks a lot like what both of those pieces describe, arrived at by a longer road: one continuous line of thought, with helpers called in for a job and let go afterward.
Anyway — I’m now down to mega basic. I wouldn’t be surprised if that says more about my own limited capacity than about the right way to build these things. Which is why I’m asking: how many agents are you running, and what are they all doing? I’d genuinely like to know.
