Multi-agent systems: behaviour between the agents

When several agents work on one task, the result depends on how they coordinate. Failure can emerge from the interaction even when every single agent behaves as designed.

Alike outside is not identical inside Alike outside is not identical inside
Alike outside is not identical inside

We test the interaction: message protocols, role boundaries, shared memory, and the paths by which one agent can instruct another. The interesting failures are usually coordination failures rather than model failures.

One recurring pattern is the confused deputy: an agent that holds permissions is asked to act by another that has none. Another is instruction propagation through shared files, where a rule written for one agent is followed by all of them.

We record the scenario, the configuration and the trace. A finding without a trace is a hypothesis, and we do not publish hypotheses as results.

Materials on this topic

Other topics

AI red teaming: what a test shows, and what it does notAI standards: where the rules actually applyAI supply chain: what enters before the model runsIncident chronicle: documented consequences, not forecastsBehavioural audit: how agent behaviour drifts over timePrompt injection: where text becomes an instructionRussian practice: who is held responsible for AIMethodology: how the Institute tests AI systemsData and benchmarks: what our conclusions rest onSociology of AI: relations between the agentsGovernment customers: how to commission a testGovernance and funding of the InstituteCases: scenarios of work and what they show

Page updated 02 November 2026 · site build 2026.11.13 · ANO «ISAIS»