Multi-Agent
Several AIs splitting up roles and passing work between them
- Multi-agent means several AIs each take on a role and pass results back and forth to finish one goal.
- Narrow a role down and instructions get short and clear. One side only researches, one only drafts, one only reviews.
- Handing off between them needs agreed rules for what gets passed and in what shape. Without rules, it's just talk going back and forth.
- Every round of back-and-forth multiplies time and cost. Without deciding when to stop, it never stops.
- More hands doesn't automatically mean more accurate. Share the same foundation and they can make the same mistake together.
Contents
1The analogy
Soccer has eleven players chasing one ball, but if everyone chased the ball, there'd be no game. There's a position that defends, one that links play through the middle, one that finishes in front of goal, and the ball moves through those positions on its way toward the net.
Multi-agent is that tactics board. Whoever gathers material hands what they found to whoever drafts, and the draft crosses over to whoever reviews. Each position does its one job, then passes to the next.
Pass the ball around beautifully and never take a shot, and no goal comes. Teams that just keep passing back and forth until time runs out are a real, common failure, which is exactly why you need to decide up front when to finish.
2In detail
Splitting positions makes instructions simple
Try to have one instruction cover everything and it grows long. Find the material, get the format right, polish the tone, check nothing's missing — pile all of that into one paragraph and something slips through.
Split positions and each instruction gets short. The research role just needs "find reliable sources and organize them with citations." The review role just needs "flag only what's missing and what numbers don't add up." Short, clear instructions produce steadier results.
Having different eyes look at the same work is another advantage of this setup. Keep the writer and the nitpicker separate, and you catch the kind of answer that would otherwise just satisfy itself and stop.
Passing the ball needs rules
The handoff is the weakest link. If the research role dumps a wall of prose, the drafting role can't tell what's a fact and what's an opinion inside it. So what goes into a handoff, and in what shape, gets decided ahead of time.
Who can hand off to whom needs deciding too. Let anyone talk to anyone, and the ball rolls anywhere, with the same thing getting repeated over and over. Narrow the handoff paths down to a few, and the flow becomes far easier to follow.
A team with a coach, and one without
One setup puts a single position in charge. It splits the work, gathers the results, and sends things back for another pass if they're short. The flow is easy to see and the results are easy to gather, but if that one position judges wrong, the whole team spins its wheels together.
There's also a setup with no one in charge, just handoffs. Whoever finishes their part passes to the next, and whoever receives it judges for themselves. It's flexible in unfamiliar situations, but harder to tell who's holding the ball at any given moment. In practice, a common compromise sets the broad shape ahead of time and only leaves a few positions room to judge on their own.
Passing without shooting means no goal
Round after round of "fix this" going back and forth forever is an extremely common failure. Reviewer flags something, drafter fixes it, reviewer flags again, and time and cost just keep piling up. Every handoff drags the whole conversation so far back into the mix, so each round gets more expensive than the last.
So you pin down how many rounds to allow, and what condition wraps things up. Rules like "stop once flags drop below two" or "wrap up with whatever you have after three rounds." Without a rule, a team can't stop itself.
Mistakes ride along with the team
If one position produces a wrong fact, the next position takes it as true and builds on it. The sentences only get smoother going forward, which makes it look even more convincing. The fact that several hands touched it is no guarantee of accuracy.
Sharing the same foundation is its own trap. The same AI with only different roles assigned tends to trip on the same confusing spot side by side, and then confirm each other and let the mistake harden. That's why numbers and facts get checked through search or a calculation tool directly, and a spot gets left where a person looks the result over before it goes out the door.
3More precisely
Multi-agent turns out to mean actually running several different models less often than you'd think. Most setups are the same model called repeatedly with a different role description each time. Giving different tools to different positions is common too — search only for the researcher, write access only for whoever saves things. Narrow each position's permissions and the damage from a wrong move narrows along with it.
The analogy breaks down in a spot. A soccer player watches the field and moves on their own judgment; the AI at each position knows nothing beyond the text handed to it. Nobody's writing down what the position next door is doing unless someone tells it to. Soccer ends when the clock runs out, but multi-agent doesn't stop on its own unless a stopping rule was written in. A soccer team also keeps the same eleven players for the whole match, while a multi-agent setup can spin up a fresh position mid-task and retire it the moment its one job is done, which makes the roster itself part of what has to be designed. Worth keeping in mind too: for a simple job, calling it once is faster and more accurate than any of this.
4Try it yourself
5Common misconceptions
It's easy to think splitting the work among several always makes it more accurate, but actually sharing the same foundation means the same mistake rides along with them, and the cost multiplies for nothing.
It's easy to think reviewing each other filters out made-up stories, but actually the next position often takes an earlier mistake as fact and polishes it into something even more convincing.
It's easy to think each position runs a different AI, but actually it's often the same AI with only its role and its tools set differently.
7One-line summary
In shortMulti-agent is a team split into positions passing the ball, and deciding when to take the shot matters just as much as splitting the roles in the first place.
Spotted an error or have a better analogy? Suggest an edit · Last updated2026-09-02