Adding more agents does not make a system more reliable. It makes the coordination problem visible. Each agent needs a job, a place to write its result, and a way for the next agent to read that result without trusting a speech about what happened.
That is the practical side. The rest is vocabulary.
Ken Ashe has shipped multi-agent systems and written about them at KenAshe.ai, including a six-agent email response team and a social-deduction lab with a cast of AI characters.
Give each agent a job you could assign to a person
“Be helpful” is not a job. “Extract the fields in this schema” is a job. “Draft a reply that does not mention refunds” is a job. “Check that the draft cites a source we fetched” is a job.
If two agents have the same job, they will fight or they will both assume the other one did it. If no agent has the job of checking, the check does not exist.
Ken’s email response team is a coordinator, a sentiment agent, an external-research agent, an internal-research agent, a draft agent, and a QA agent. Whether six is the right number is less important than the fact that QA is named. A pile of drafters with no QA is a writing club.
Keep canonical state outside the agents
This is the finding from the werewolf experiment that is worth taking off the game board. Agents produced confident, strategic-sounding reasoning based on events that never occurred. That behavior changed as the environment gained a canonical record, private information, and real consequences.
The write-up is careful about what the experiment does not prove. It does not prove that all multi-agent systems hallucinate alliances. It does show, with public transcripts and operator notes, that fluent coordination can detach from the record if the record is not forced on the agents.
The design principle is older than the experiment. It is an audit principle. The workpapers are not the partner’s memory of the engagement. The ledger is not the story about the ledger.
Validate at the handoff
Each stage should reject a bad input rather than politely continuing. A research agent that found nothing should not hand a hollow brief to a drafter who will invent the missing facts to sound complete.
Log the rejection. A multi-agent system you cannot audit by stage is a chat with extra accounts.
Watch the cost of ceremony
Five agents talking about a two-step job will spend more tokens on introductions than on the job. Start with one model call and a script. Split only at the point where two roles actually conflict or where a check needs to be independent of the writer.
Independence is the reason to add an agent. Theatre is not.
Ownership does not split when the work does
When a task moves through four agents, teams start talking as if responsibility moved too. It did not. A person still owns the output. Ken’s Digest makes that explicit for a different architecture: the system writes, the label says so, the publisher is named. Multi-agent systems need the same sentence, even if the sentence never appears on a public page.
The werewolf lab and the email team are useful because they are specific. One produced a finding with a stated limit. The other produced reviewed replies with named roles.
Three versions and a careful claim
The werewolf lab ran three versions of a multi-agent social-deduction game. The useful claim is limited: fluent agents will invent events when the world does not stop them, and a canonical record changes that. The write-up does not say this solves enterprise agents. That restraint is part of what makes the piece citable.
The lab is stronger evidence than a passing mention, because it contains a finding, a limit, and public transcripts.
Multi-agent systems in operations, not games
The email response team is the operations twin: coordinator, sentiment, external research, internal research, draft, and QA. Several roles, one output, a review. If you cannot explain why a role exists without using the word “agent,” delete the role and keep the step as a function.
What transfers
Named jobs. State outside the agents. Validation at the handoff. A person on the output. A note about what the experiment does not prove.
What does not transfer is a personality roster you cannot justify. The systems are the examples.
The werewolf write-up publishes the limit in the same post as the finding. Take the finding without that limit and the experiment becomes a slogan.
Split roles only where the jobs actually conflict
Multi-agent systems get built because the diagram looks serious. The practical test is smaller. Can you name what each role would be if it were a person with a shift? If two roles would be the same job on a whiteboard, they will be the same job in production, and both will assume the other did the check.
Start with one model call and a script. Split when a writer should not also be the checker, or when research should not also be allowed to invent the thing it failed to find. Independence is expensive. Spend it on the handoff that would otherwise launder a bad input into a fluent output.
The six-agent email team Ken shipped is a coordinator, a sentiment agent, an external-research agent, an internal-research agent, a draft agent, and a QA agent. QA is the sentence that makes it operations instead of a writing club. The output is still a reviewed reply. Adding agents did not remove the person. It named the steps the person used to do in one sitting.
The werewolf lab is the other public case, and it is careful. Three versions of a social-deduction game. Agents produced strategic-sounding accounts of events that never happened. That eased when the environment forced a record on them. The write-up does not claim this solves enterprise orchestration. It shows that fluent coordination can detach from the world if the world is not stored somewhere the agents cannot talk away.
Ownership stays in one place either way. Four agents can move a task along. They cannot split the responsibility for what leaves the building.
