Consensus Mechanism
Run the same question past five agents and you may get five answers. A consensus mechanism is the rule that turns them into one — a vote, a few rounds of argument, or a designated judge that reads every position and decides.
The common forms
Four patterns cover most systems in use:
- Voting — agents answer independently and the majority wins. Cheap, parallel, and effective where answers are short and comparable.
- Debate — agents see each other’s reasoning and revise across two or three rounds, so a well-supported minority answer can pull the group over.
- Aggregation by a judge — one agent, often with a different prompt or a stronger model, weighs the submissions and issues the final answer.
- Verification — where a claim can be checked by a tool or a test, the check settles it and the vote becomes a tie-breaker.
Why it earns its cost
A single agent’s confident mistake has nothing to push back on it; several agents, if their errors are independent, cancel out. That is the whole bet, and also where it fails. Agents built on the same base model with the same prompt make correlated errors, so they agree on being wrong and the mechanism launders that into a confident result. Debate has its own weakness: agents defer to whichever position was stated most assertively. Frameworks such as AutoGen make these patterns easy to assemble, which leaves the discipline as the hard part — every extra voice multiplies tokens and latency, so the rule has to be worth the spend it commits the system to.
A consensus mechanism is the rule by which several agents working on the same question settle on one answer — majority voting, a few rounds of debate, or a judge agent that weighs each position. It exists to cancel the confident errors a single agent has nothing to check, and it fails when the agents share a base model and prompt: their mistakes correlate, and agreement stops being evidence.
When a committee has to decide something, there is always a rule for finishing: a show of hands, a discussion until minds change, or a chair who listens and calls it. Without a rule, the meeting never ends.
A consensus mechanism is that rule for a group of AI agents working on the same problem. They each produce an answer, then vote, argue it out for a round or two, or hand everything to one agent that decides.
The reason to bother is simple: one program can be confidently wrong with nothing to contradict it. Several can catch each other — unless they all think alike, in which case they just agree on the same mistake.