Home / Several agents and adversarial checking

Multiplying agents can multiply the error

A study testing five agent architectures on 260 configurations measures an error rate amplified to around seventeen times that of a single agent when independent agents work without a coordinator validating their outputs, against around four times when a coordinator does validate them, and on a strictly sequential task, all four multi-agent architectures tested fall behind a single agent.

A study published on arXiv in December 2025, updated in April 2026, tested five agent architectures on 260 configurations, six benchmarks and three different model families. It measures an error amplification rate at the execution trace level, called A trace in the study, which compares the number of errors produced by a multi-agent system to the number of errors from a single agent on the same task; this figure appears in table 5 of the full text, not on the summary page alone. When independent agents each work on their own share without a coordinator validating their outputs, this rate reaches around seventeen times that of a single agent.

The coordinator that validates changes everything

The same study measures a much lower amplification, around four times, for the centralised architecture, where a coordinator reviews and validates each agent's output before it is passed upward. The authors explain this gap by a deliberate bottleneck: the coordinator acts as a mandatory checkpoint that can intercept an error before it spreads. The independent architecture has no equivalent checkpoint, each agent produces its result without ever getting the chance to check it against another's.

Study, 260 configurations, six benchmarks, three model families
Independent architecture, no validating coordinator: A trace = 17,2
Centralised architecture, validating coordinator: A trace = 4,4
Strict sequential task (PlanCraft): all four multi-agent architectures
fall behind the single agent, by -39 to -70 per cent

A strict sequential task levels everyone down

On a task that imposes a strict order of steps, called PlanCraft in the study, all four multi-agent architectures tested fall behind a single agent, by between 39 and 70 per cent depending on the architecture chosen. Adding agents does not automatically improve a result: on this type of task, the opposite happens, whatever way the agents are organised among themselves.

At the time this course was written, this study remains an arXiv preprint, at its third public version: no mention of a final publication in a peer-reviewed journal was found in the text itself, so this point remains to be verified rather than asserted. The previous lesson details what an adversarial verifier needs to receive to judge a piece of work; the next lesson shows another place where several agents multiply the same weakness rather than correcting it, a false claim slipped into the framing given to each of them.