Role drift (multi-agent systems)

Role drift is the gradual erosion of an agent’s assigned role over a long run: the agent keeps returning well-formed output in the right register while the remit it is actually working to widens, narrows, or slides onto a neighboring agent’s territory. No individual turn in the run reads as a violation. The deviation becomes visible only when a late output is held against the brief the agent opened with, which is why role drift survives a trace review that catches abrupt failures.

Our account of how multi-agent systems fail treats role drift as the slow form of a specification fault, and the empirical taxonomy behind that page annotates the end state rather than the slope. In MAST’s distribution across the 1,642 traces its LLM-as-judge pipeline labeled, disobeying the role specification accounts for 1.5% of failures and task derailment for 7.40% (arXiv:2503.13657, v3, NeurIPS 2025 spotlight poster). Both figures are descriptive shares of a machine-labeled corpus, so read them as a picture of what got annotated rather than as incidence rates for your own system. Downstream the damage is ordinary: a widened remit produces output the next agent treats as authoritative, which is error propagation with nothing in the output to flag it, and finding the turn where the remit first moved is a backward attribution problem. Compare a late output against the original brief, never against the turn before it.

How to measure role drift

Because drift is a slope rather than a state, a single end-of-run check cannot separate it from one lapse, since both leave exactly one bad output behind. The quantity to record is role adherence indexed by turn, scored against the same written contract at every point in the run, so that a declining series stays distinguishable from a flat series with one dip in it. Each turn either adheres or it doesn’t, so every bucket you aggregate into is a proportion on a small denominator and needs an interval before anyone reads a trend off it.

Write the role as something checkable before you measure anything: what this agent may produce, what it must decline, and which decisions belong to another agent. Score every turn of a long run against that contract as adhering or not adhering, then group the scores into turn buckets, turns 1 through 5, 6 through 10, and onward at whatever width the run length supports. Drift is the decline in adherence across those buckets. Report each bucket’s adherence rate with its own confidence interval rather than a fitted line, because buckets built on ten or twenty turns produce intervals that overlap far more than a plotted slope suggests. The pass-rate interval calculator takes k of n directly, and the interval is compulsory here for the same reason it is compulsory behind any eval confidence interval.

The closest measured analog sits in the single-model case. Li and colleagues benchmarked instruction stability through self-chats between two instructed chatbots and found instruction drift setting in within eight rounds of conversation on LLaMA2-chat-70B and GPT-3.5, tracing it to attention decay over long exchanges (arXiv preprint 2402.10962, last revised July 2024). Those are 2023-era models in a two-party dialog rather than an agent graph, so the eight-round figure carries over as a reason to bucket by turn index at all rather than as a rate you should expect. A system prompt that loses its grip within single-digit rounds says nothing encouraging about a worker forty tool calls into a run.

Instrument at the point where the role contract is re-read, which in practice means the orchestrator’s dispatch to each worker, or the step in a sequential or supervised topology where one agent inherits an accumulated transcript.

Role drift vs inter-agent misalignment

Inter-agent misalignment names a defect in the exchange between agents at one moment: a constraint dropped from a message, a peer’s correct signal discarded, a sub-dialogue restarted, whatever goal each agent was honestly pursuing at the time. It is a property of the coordination surface between them, and role drift is a property of one agent measured across time.

Neither of the two readings constrains the other. A system can hold perfect role adherence and still be badly misaligned, when every agent works exactly to its brief and the message that would have carried the deciding constraint never arrives. It can also drift severely with no misalignment anywhere, when one agent widens its remit into territory nobody else is working in and the widening goes unchallenged because no other agent has a stake in it. Look for misalignment in what passed between two agents, and for drift in what one agent was doing at turn five against what it is doing at turn fifty.

Role drift vs rubric drift

Rubric drift is the judge-side term built on the same word. There the effective standard an LLM judge applies moves across a run or between prompt versions, so an output that scored a pass early would score a fail late without having changed at all. Role drift moves what is being scored, and rubric drift moves the scale it is scored on. Both surface as a score series that trends downward across a long run, and they demand opposite repairs, since re-anchoring a worker’s role does nothing at all about a judge whose threshold has crept.

The separation test is cheap, and it runs on outputs you already have. Re-score a batch of early outputs using the judge exactly as it stands at the end of the run. If those old outputs now score differently, the scale moved, and the trend you’d been reading as role drift belongs at least partly to the judge.

Role drift already means something in organizations

Organizational and staffing writing uses role drift for a person’s job quietly expanding past its written description, usually in company with role creep and scope creep, and that sense is older and better settled than this one. The remedy there is governance: renegotiate the description, or staff the gap the drift had been covering. That remedy does not port to an agent. An agent has no interests to renegotiate and no workload to relieve, and rewriting a prompt after a run has finished repairs nothing about the run that already happened.

The mechanism underneath the agent version is context accumulation, where the agent conditions on a transcript that grows away from the instruction that opened it until that instruction is a shrinking fraction of what the model attends to. Replace the governance fix with a mechanical one: re-anchor the role contract inside the run at a cadence you chose deliberately, then measure adherence over the turns following each re-anchor to find out whether the re-anchor held.

Role drift erodes what an agent takes its job to be, and context handoff loss erodes what it was told. Both are slow, and neither leaves a turn you can point at. The fuller account of how specification faults travel through a topology, and which structural levers bound how far they get, sits in our multi-agent failure taxonomy, and the sibling terms this one is scored against are laid out in a term-by-term index of the family. A team whose runs finish inside half a dozen turns has no slope to measure, and for them per-turn scoring may not be worth the labeling it costs.