Motivation¶
Silent
"None of these [cognitive failures] produce a latency spike, an exception, or an error code. They are operationally silent." — Silent Failures, Loud Consequences
Loud
"Five agents do not catch the error; they ratify it. And the system's output arrives with the strongest possible success signal a debate can produce: everyone agreed." — Silent Failures, Loud Consequences
The problem¶
Multi-agent systems now decide: which peer to ask, whether the evidence is enough, when the work is done. Monitoring still measures execution: latency, errors, cost.
A cognitive failure is an agent doing a plausible thing for a broken reason. It verifies without examining, agrees without reconciling, or asserts without knowing. The trace shows success, so no existing signal fires.
The consequences¶
Being invisible does not make these failures minor. Measured against a ground truth, debate accuracy runs at 0.99 with no cognitive failure present, 0.81 with two categories in the trace, and 0.16 at three — approximately the rate expected from guessing. Verification collapse alone costs 87 accuracy points.
The failure modes also compound in a way that inverts the usual benefit of redundancy. Where the mechanism that would normally catch an error is itself compromised — dissent suppressed, or verification no longer verifying — additional agents do not correct the error. Under verification collapse, five agents reach unanimous agreement on a wrong answer 75% of the time.
This is explored in more detail in the tech blog.
From diagnosis to response¶
Cognitive failures can be split into two classes:
- Structural failures trace back to a specific design decision, so they can be fixed once identified.
- Emergent failures arise at runtime from the task, the context or the model itself, so no design change removes them and they have to be caught while the system runs.
Each class of failure requires its own answer, as summarized in the table below:
| Kind | Cause | Answer |
|---|---|---|
| Structural | A specific design decision (e.g. a crash-only test gate) | Name it, reproduce it, design it away with MAS-Lab |
| Emergent | The task, the context, or intrinsic model behaviour | Detect and correct at runtime |
IoC addresses the structural class with MAS-Lab, an open-source framework allowing to iterate on the design of a MAS and quantifying the improvements (or regressions) from one version to another.
The emergent class of failures is addressed by IoC with Cognition Protocols for runtime coordination, detection built on the SCOPE observability substrate, and potentially MAS-Lab to validate each claim.
Source¶
Silent Failures, Loud Consequences: Cognitive Challenges in Multi-Agent Systems., September 2026.