Context handoff loss (agent to agent)

Context handoff loss is the context an agent needed, held by the agent working before it, and never received when the work crossed between them. The taxonomy term underneath it is MAST’s loss of conversation history, defined there as “Unexpected context truncation, disregarding recent interaction history and reverting to an antecedent conversational state” (Cemri et al., 2025, arXiv:2503.13657, NeurIPS 2025 spotlight poster, 14 modes over 1,642 annotated traces). This lane’s name puts the loss at the boundary, which is where it stops being one agent’s memory problem. The receiving agent gets a shorter brief than the sender was working from, treats it as the whole brief, and fills the gap with something plausible. Nothing in the run errors, which places this in the same family as an agent reporting a success it never earned.

Loss at an edge is the cheapest place to intercept error propagation, since a constraint that never crosses the boundary never becomes a downstream premise, and the lever that bounds it there is containment at that hop. Which edge dropped the item is a backward question for failure attribution, and how far the resulting error traveled afterwards is propagation radius. Wiring changes the exposure: a sequential pipeline offers one long chain of edges to lose things on, while a supervisor topology routes each dispatch back through an agent that still holds the original brief. Instrument the edge, or you will spend the postmortem diagnosing a receiver for a fault it inherited.

Context handoff loss vs information withholding

Information withholding is MAST’s own mode, defined in the paper as “Failure to share or communicate important data or insights that an agent possess and could impact decision-making of other agents if shared.” The sending agent holds the detail and does not pass it on. Context handoff loss names the case where the sender did commit the context to the handoff and the receiver still did not get it, or got a version with the load-bearing part gone.

Where the defect lives is what separates them, and each location points at a repair the other does not need. Withholding is a policy defect in the sender: its instructions, its summarization prompt, or its working notion of relevance decided the detail was not worth passing, so the repair sits upstream, in what the sender is required to emit. Handoff loss is a transport or representation defect, produced by a window eviction, a compaction pass, a serialization that dropped a field, or a summary written to a length budget instead of a contract. You repair that one at the boundary, by typing the message and validating it on receipt. Confusing them sends you to rewrite a prompt when the message format was at fault, or to harden a schema when the sender was never told to include the field. Both belong in the same per-edge audit, since one trace can carry an instance of each.

How to measure context handoff loss

Measured, it is a rate on one edge: over repeated runs of the same workflow, the share of handoffs where at least one required item was absent from what the receiver got. The denominator is the number of handoffs observed on that edge, and the quantity is a proportion, so it carries an interval like any other pass-style rate. The harder half is the numerator, because a loss rate has no meaning until the edge carries a contract naming what the receiver must have. Writing that contract is a precondition of the measurement rather than a step that follows it.

Write the contract for one edge first: the items the receiving agent must have to act correctly, expressed as claims rather than as fields. Typical entries are a budget ceiling the user stated three turns back, a rejected first design, and one data source that came back empty. Then run the workflow many times and compare each message arriving at that edge against the contract. Loss rate is the count of handoffs missing at least one contract item over the count of handoffs observed. Report it with a confidence interval, on the reasoning behind any other eval confidence interval; the Wilson and Clopper-Pearson calculator does the arithmetic from those two counts.

A sharper second reading comes from fault injection.

Delete one known contract item at the edge deliberately, then count how often a downstream agent notices before the run ends. The detection share reports whether the loss is survivable, since an agent that spots the gap and asks behaves differently from one that fills it silently. The flagship profiler is designed to seed exactly that kind of single-item omission at a chosen edge, replay the run across many trials, and report the detected share with a bootstrap interval beside it. The profiler is unshipped, so the detection share on your own edges is yours to measure; for what a measured run looks like once it is written up and taken apart, the studies on the research track are the worked examples. Instrument where the message is serialized for the next agent, ahead of any compaction step, since a log captured afterwards records the loss as the sender’s own output.

Truncated tool output just upstream of an edge produces the same signature, so rule it out before blaming the handoff.

Context handoff loss vs conversation reset

A conversation reset discards state a session was holding, so the agent that resumes works from an earlier point in its own history. Context handoff loss is about what crosses an edge between two agents, and it can fire on the first handoff of a run that never reset anything.

The two do coincide at one point in a run, and the overlap distorts the number. A reset upstream of a handoff guarantees loss on that handoff, since the sender can only forward what it still holds, so a system that resets often shows a loss rate the edge itself did not earn. Reading them apart takes a trace carrying the sender’s state before the handoff alongside the receiver’s inbound message, and a trace of final outputs will separate neither.

Handoff communication failure in medicine and in agent systems

Medicine has owned this word for decades, and the collision is worth conceding plainly. AHRQ’s patient-safety primer defines a handoff as “the process of one healthcare provider updating another of the status of one or more patients for the purpose of handing over responsibility for their care” (AHRQ PSNet, primer, last reviewed 2024-06). Anyone searching handoff communication failure lands mostly on clinical work, appropriately so. Clinical practice also arrived at something worth borrowing: structured handover protocols exist to stop the same loss this page describes, and the I-PASS mnemonic closes on “Synthesis by receiver: an opportunity for the receiver to ask questions and confirm the plan of care.”

That closing step is where the borrowing stops being free. Synthesis by receiver works because a clinician hearing an incomplete report can feel the shape of the gap and ask about it. A receiving agent usually cannot, because it holds no record of what it was not sent, so an absent constraint reads to it as a constraint absent from the world. Copying the protocol into a prompt gives you the ceremony and none of the check. The version that transfers makes the synthesis machine-checkable, having the receiver restate the items it believes it holds and stopping the handoff when that restatement misses one.

An edge that loses context and catches it is a different system from an edge that loses context and ships, which is why the loss rate and the detection share belong side by side in the same report. The taxonomy this term sits inside, and the containment lever that bounds each mode once it fires, are worked through in our map of multi-agent failure modes, every mode in this family, named gives each neighboring mode its own entry, and the failure-side definition of a multi-agent system accounts for why these edges exist at all. If you can only afford one move on one edge, then write the contract before the instrument, because a loss rate with no contract behind it counts against a standard nobody set.