Information withholding (multi-agent)

Information withholding is the failure mode in which one agent holds data or an insight that would change what a peer does next, and the message it sends leaves that data out. The receiving agent has no way to learn the gap exists, so it fills the space with a plausible guess and carries on as though the brief were complete. MAST, the empirical taxonomy of multi-agent LLM failures, catalogs it as mode FM-2.4 in the inter-agent misalignment family (Cemri et al., Why Do Multi-Agent LLM Systems Fail?, arXiv:2503.13657, NeurIPS 2025 spotlight poster, as of 2026-08). Its description there is a failure to share data or insights an agent possesses that could change what another agent decides.

The other modes in that family leave a mark on the transcript. Withholding leaves none, because the defect is a counterfactual: something the sender had and did not pass on, which a message log by itself can never show. Scoring it therefore takes two records, the sender’s available context at the moment of the handoff and the payload that actually left, and the comparison between them is the measurement. That is also the likeliest reason it is the rarest mode in the taxonomy’s own distribution, at 0.85% of annotated failures, the smallest share of the fourteen. Treat that decimal as provisional rather than settled. The paper attaches no interval to it, and the two revisions do not count the same objects: the April 2025 version analyzed over 200 tasks, while the October 2025 one reports 1,642 annotated traces. Tasks and traces are different units, so quote the revision alongside any percentage lifted from this paper.

It sits under inter-agent misalignment as one specific shape of defective message, and it feeds error propagation the moment the receiver’s guess hardens into a shared premise that later agents reason from. In a post-mortem it is among the hardest faults to attribute, because the earliest corrupted node looks well formed on both sides of the boundary, and the missing item was never in the trace to begin with.

A run can end with a correct-looking answer and still contain a withheld fact, so the unit you measure is the handoff rather than the final result.

How to measure information withholding

Write down, per handoff, the items the receiving agent needs in order to act correctly: the identifiers, the constraints, the caveat attached to an earlier result. Instrument two surfaces at that boundary, the context available to the sender and the payload it emits, then count the handoffs where a required item was present upstream and absent downstream. The withholding rate is that count divided by the handoffs where the item was available to withhold in the first place, which keeps the denominator honest in the way answer coverage does for a pass rate. Because the result is a proportion, report it with an interval rather than as a lone percentage. The interval is what tells you whether a change between two builds is real movement or sampling noise, and the pass-rate interval calculator produces one from the two counts.

The controlled version of the same measurement is cleaner. Plant a fact that only one agent can see, make it necessary for the downstream step to succeed, and observe whether it crosses the boundary. That is ordinary fault injection with the fault placed in an agent’s knowledge instead of in its tools.

Information withholding vs context handoff loss

Context handoff loss names what the receiving agent fails to get because the transfer itself dropped it: a summarization step that compressed away a constraint, a window that rolled, a serializer that quietly never carried the field. Information withholding names context the sender had available and did not select for the message, with the transport working exactly as designed.

A real system can show either one on its own. It can lose context at every handoff while withholding nothing, when the sender assembles a complete brief and the pipe truncates it. It can withhold at every handoff while losing nothing, when a faithful transport carries a message the sender composed without the caveat sitting in its own scratchpad. Their repairs live in different places, which is the practical reason to keep the two apart: handoff loss is fixed in the transport and the summarizer, withholding in whatever decides what goes into the message. Measuring only this one leaves what gets dropped in transit unscored, and the pair together accounts for the brief the receiver ended up holding.

Information withholding vs knowledge hiding

Knowledge hiding is the established construct in organizational behavior, and it is a well-built one. Connelly, Zweig, Webster and Trougakos defined it in 2012 as an intentional attempt by an individual to withhold or conceal knowledge that another person has requested (Journal of Organizational Behavior, 33(1), 64–88). They separated three strategies for doing it: playing dumb, evasive hiding, and rationalized hiding.

Two premises in that definition fail to survive the port to agents. Start with the request: in a pipeline the downstream agent usually never asks, since nothing in its input tells it the fact exists. Intent is the other, and an eval has no instrument that reaches it, because a trace shows an omission and doesn’t say why it happened. So the version an eval can defend drops the motive and keeps the boundary, scoring an item that was available to the sender, needed by the receiver, and absent from the message. The behavioral surfaces do still rhyme, and an agent reporting that it lacks a detail already sitting in its own context looks a great deal like playing dumb. The evidence supports one label: an omission counted at the handoff, with its denominator attached.

A typed handoff contract turns a withheld field into a rejected message

Prevention runs through whichever component composes the outgoing message. Give a handoff a declared schema whose required fields include the provenance and the caveats, and an omission stops being an invisible choice by the sender. It becomes a validation error at the boundary, which is an event a containment measurement can score. The weaker form of the same lever is a checklist in the sender’s prompt, and it decays over a long run the way any prompt-level constraint decays once context accumulates. Topology sets how much enforcement costs: a hub-and-spoke design routes every message through one composer, so a single contract covers the whole fan-out, while a mesh of peers needs the contract on every edge. Choosing the coordination pattern therefore also chooses how many seams you have to hold.

Two things can leave a downstream agent reasoning from a smaller brief than the one that existed upstream, and this is the one where the transport worked. Both cost less to catch at the seam than to unwind from a finished run. The other seven terms in this family sit in the term-by-term index for multi-agent failures, and the fuller account of how a coordination fault becomes a system failure, with the containment lever that bounds each one, lives in the multi-agent failure taxonomy.