Inter-agent misalignment

Inter-agent misalignment is the failure class that lives in the traffic between agents: ineffective communication, conflicting behavior, and gradual derailment from the objective the agents were jointly handed. MAST, the empirical taxonomy that named it, makes it one of three top-level categories (Cemri et al., Why Do Multi-Agent LLM Systems Fail?, arXiv:2503.13657, NeurIPS 2025 spotlight poster, as of 2026-08). It sits between the specification faults wired in before a run starts and the verification faults that let a wrong answer out at the end. What puts a mode in this category is where its damage lands, on the coordination surface between agents, whatever produced it.

Inter-agent misalignment placed second of the three categories in that analysis, just under a third of annotated failures, which is why the failure-mode pillar publishes the ordering and leaves the decimals alone. MAST annotated other people’s systems, so the prevalence you inherit is a prior about how popular frameworks wire agents together, and your own rate stays an empirical question about your own graph.

Placed in the lane’s vocabulary, a misaligned message is the trigger, error propagation is what happens to it next, and propagation radius is how far it traveled before something caught it. Working backward from a wrong final answer to the exchange that started it is failure attribution. The containment lever for this category sits at the edge between agents rather than downstream of it, because a retry on the receiving agent re-reads the same defective message.

The six modes the category covers

Six annotated failure modes sit under inter-agent misalignment, and what they share is a landing site rather than a cause: every one of them shows up on the coordination surface between agents, and two of the six start inside a single agent’s own reasoning before they get there.

Reasoning-action mismatch and failure to ask for clarification are the two most frequently annotated modes in MAST’s category. Both are failures of an agent’s own discipline, so a typed handoff between agents will not catch either. All six carry their own entries here, because teams hit and debug them separately: those two, plus task derailment, conversation reset, information withholding and ignored agent input. Adjacent shapes that the taxonomy files elsewhere have entries too. Role drift is derailment applied to an agent’s remit instead of the objective, and context handoff loss covers context that goes missing in transit rather than being held back. The term-by-term index for this vocabulary routes to every term in the family that carries an entry.

The category is a label applied to a trace, so every figure attached to it is an annotation rate and inherits whatever the annotators agreed on. MAST built its taxonomy from 150 execution traces read by expert human annotators (kappa 0.88), then scaled the labeling across seven frameworks with an LLM-as-judge pipeline. The per-category distribution it publishes is attributed in the Figure 1 caption to an analysis of 1,642 execution traces, and carries no confidence interval. That denominator is worth pinning to a version. The April 2025 revision described an analysis across seven frameworks and over 200 tasks, and the October 2025 one reports 1,642 traces, so a percentage quoted from this paper means little without the version it came from. Report the agreement statistic beside any prevalence you quote, because a prevalence computed from labels nobody agrees on measures the labelers.

How to measure inter-agent misalignment

You measure it by labeling rather than by computing. Sample execution traces from the system you actually run, read each one forward to the first point where a message between agents was defective, and record the mode that fits. The denominator is traces sampled, since one trace can carry several defects and counting messages instead rewards a terse topology. Have a second annotator label the same sample and publish the agreement alongside the rate. The labeled share is a proportion, so it wants an interval, and the pass-rate confidence interval calculator takes k of n and returns a Wilson bound.

Instrument at the handoff itself, which in most stacks means the orchestrator step that forwards a worker’s return, or the write into shared state. A labeler needs the message as sent and the receiver’s next action in one record before a mismatch can be called at all. How much of that you can see is a topology question, and the orchestration patterns differ in whether messages route through a hub you can log or move peer to peer.

Inter-agent misalignment vs AI alignment

Alignment in the AI-safety literature is about targets rather than channels. Ngo, Chan and Mindermann state the concern directly, arguing that advanced systems “could learn to pursue goals that are in conflict (i.e. misaligned) with human interests” (arXiv:2209.00626). Misalignment there is a relation between a system and people. Inter-agent misalignment is a relation between two components of one system. Every agent in a misaligned trace can be pursuing exactly the goal you gave it, honestly and competently, while the run still returns a wrong answer because one message left out a constraint.

Safety alignment and coordination can each hold while the other fails. Two agents perfectly aligned in the safety sense will still derail when the handoff between them drops a constraint, since coordination is a property of the channel and no amount of good intent repairs it. A single agent misaligned in the safety sense produces harmful output with no coordination anywhere in the picture, and putting a second agent beside it supplies a peer to be persuaded.

The collision is real, and the safety literature holds the older claim on the bare word. So the compound keeps its qualifier: write it in full on first use, and drop to the short form only where the sentence around it has already established that the subject is a topology.

Inter-agent misalignment vs role drift

Role drift happens inside one agent over time: the brief holds for the opening turns, erodes as context accumulates, and the agent widens what it treats as its job while returning well-formed output every turn. Inter-agent misalignment is a property of an exchange between two agents at a point in time, and the evidence differs accordingly. Drift is visible by comparing a late output against the original specification, since no individual turn reads as wrong on its own. Misalignment is visible in the transcript, as a message that says less than the receiver needed or an action that contradicts the plan just stated.

They also compound, which is why a team that instruments for one keeps finding the other. A drifted agent goes on sending well-formed messages, and its peers coordinate correctly around a role nobody assigned, so every edge in the transcript looks clean while the output answers a question you did not ask. Label the two separately in your annotation scheme, because the repairs do not overlap: drift is answered by re-asserting the brief against the original spec at intervals, and misalignment by a typed, verified handoff at the edge.

A coordination defect either stays inside the system or leaves as a certified answer, and task verification failure is the category that decides which. That is how a topology with disciplined handoffs and no exit gate still ships wrong work to a user. The fuller account, each category mapped to how it propagates and which lever bounds its reach, lives in multi-agent LLM failure modes; the structural vocabulary underneath it, agent and orchestration and shared state and topology, is set out in multi-agent systems defined by how they fail.