How do swarms of AI agents actually work together?
People describe frontier labs running swarms of thousands of agents on huge projects. I already struggle to get two or three agents to cooperate reliably. How are large groups of agents coordinated in practice, and what do you actually gain by running that many?
Nobody outside those labs can show you how a thousand-agent setup is run, so treat those descriptions as claims, not a recipe. What is public, and what works at small scale, is the same pattern: one coordinator splits the work, many workers each do one isolated task, and results come back to be checked and merged.
The gain comes from independence, not from agents talking to each other. Agents that freely message one another tend to drift off course. Tasks that do not share files or state can run in parallel safely; tasks that depend on each other still have to run in order, no matter how many agents you have.
If two or three agents already fail for you, the usual cause is overlap: they edit the same files, or each one lacks the context the others built up. Give every worker a written task with its own files and a clear definition of done, and have a single reviewer check each result before the next step.
Scale up only after that loop is reliable at three agents. More agents multiply both throughput and mistakes, and the review step is what keeps the second from outrunning the first.
Listings mentioned
- Dispatching parallel agents · skill by obraSends two or more independent tasks to separate agents that share no state.
- Subagent driven development · skill by obraRuns a plan task by task with a fresh subagent each time and a review between tasks.
Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.