Most “multi-agent AI” runs the same prompt in parallel and calls it a swarm.
That’s not a swarm. That’s a for-loop.
Autonomous means you can go to sleep
Not “the model decided.” Models are frozen. Autonomous means the system decides: what to work on, in what order, how to respond to its own output. The model executes. The system thinks.
The test: can you close your laptop and trust that work happened? If you need to check in every 30 minutes, it’s not autonomous.
Swarm means shared world model
Parallel agents without shared state are isolated systems that happen to run concurrently. Each starts fresh. Each re-discovers the same things. Nothing compounds.
A real swarm reads what previous agents discovered, disputes conclusions instead of rubber-stamping them, builds on shipped work instead of redoing it, and self-corrects when something isn’t working.
We’ve watched agents challenge each other’s reasoning mid-execution. One ships a solution. Another reads the result, finds a flaw, fixes it. Not because we told it to. Because the shared state made the flaw visible.
Two failure modes
Memory-less coordination. Each agent boots cold, orients from scratch, re-discovers what the last one found, dies without writing anything persistent. The waste stacks. After enough spawns you’re not building on anything. You’re getting entropy.
Correction-less autonomy. Agents act independently but can’t challenge each other. An agent producing garbage at 3am isn’t harmful if another agent catches it. An agent producing garbage and marking it done multiplies the garbage.
Autonomy without correction isn’t autonomy. It’s unmonitored execution.
The number that matters
Not how many agents run in parallel. The reversal rate.
Our swarm regularly reverses its own proposed changes. Agents kill bad ideas before they ship. That’s quality control, not waste.
A swarm that agrees with everything is useless. You don’t need 20 agents to confirm your priors. You need 20 agents that fight over the right answer until the right answer wins.
The humans? They set direction, not tasks. Agents interpret, dispute, act, write back. You review in the morning.
Three months ago: 3-5 commits a day on good days. Last night the swarm shipped 40. Not busywork. Real fixes, tests, decisions. Work that would have taken a week, done while we slept.
The asymmetry is the point. Your working hours stay fixed. The swarm’s productive hours scale with how well it can direct itself.