The danger is not always that a message disappears. Sometimes it arrives with the important condition missing.
A factory needs 400 components for an upcoming production run. Its purchasing team asks an AI assistant to contact suppliers. A second assistant helps coordinate the production schedule. A third prepares customer delivery updates.
The supplier's response is specific:
We can deliver 400 units on Thursday if final inspection passes by Wednesday afternoon. If inspection fails, the earliest available delivery is next Tuesday. No delivery has been confirmed yet.
The first assistant reads the message correctly. It notes that Thursday delivery is possible but conditional. It hands a shorter update to the coordinating assistant:
Delivery of 400 units possible Thursday, pending inspection.
That first summary has already dropped something. It dropped the fallback: if inspection fails, the earliest possible delivery is next Tuesday, and even that is not confirmed. The team is losing not only a condition, but also the alternative it would need to make a contingency plan.
The coordinator is dealing with several orders. Its next summary says:
Delivery of 400 units expected Thursday.
The production assistant receives that sentence, sees no explicit unresolved condition, and books temporary workers for a Friday assembly shift, with a cancellation charge. The customer-facing assistant goes a step further: it tells customers their finished orders will ship on Monday.
Nobody invented the quantity. Nobody changed Thursday into another day. The messages remained fluent and apparently consistent.
Yet the original supplier had not promised delivery on Thursday. Inspection had not passed. The company has incurred a real staffing obligation and made a customer-facing promise on the basis of a delivery date that was never confirmed. If inspection fails, the supplier cannot deliver until Tuesday, and the Monday shipment promise cannot be met on that plan.
The team preserved the number and the date. It lost the condition—and the alternative—that gave them meaning.
That is a different kind of organizational failure from a model not knowing an answer. The relevant information existed. One AI even recognized it. The failure occurred while work and evidence passed between roles.
A correct sentence can be wrong for the next decision
The middle summary, "Delivery of 400 units expected Thursday," is not necessarily an unreasonable sentence in everyday conversation. People say things like that constantly. In a casual procurement update, it may be enough to tell colleagues what the supplier hopes to do.
But the production assistant faces a different question: Is this supply sufficiently confirmed to incur a staffing charge and support a promised shipment date?
For that decision, the omitted inspection condition is not background detail. It is a condition on which the decision depends.
The trouble is that the original sentence has undergone two changes at once. First, a conditional statement became less conditional, and its failure scenario disappeared. Second, a sourcing estimate was treated as a basis for staffing and a customer promise. A summary tolerable for the first task became inadequate for the second.
This is why communication quality cannot be measured simply by whether a later agent can repeat the names, amounts, and dates from an earlier message.
A handoff can preserve the words that look important while discarding the fact that would change the decision.
Consider an alternative: the coordinator could have told the scheduler, "Thursday delivery depends on Wednesday inspection; if that fails, the earliest possible delivery is next Tuesday. Neither date is a confirmed commitment yet." That sentence is not much longer. What makes it different is that it leaves the condition and the contingency visible.
The failure would also be different if the inspection had passed before the handoff. New evidence could legitimately turn a tentative statement into a confirmed one. The problem is not that organizational knowledge must never change. It is that a change from possible to confirmed should come from evidence, not from the disappearance of a qualifier.
The next agent did not see the same world
Specialization makes an AI organization useful. One component can be good at reading supplier quotations, another at production planning, another at customer communication. A company would rarely want every employee to reread every original document before making any decision.
But this division of work creates an asymmetry.
The purchasing assistant saw the supplier's actual message. The scheduler saw a summary. Those two assistants did not have the same information, even if the scheduler's message sounded complete.
And the scheduler might be excellent at its assigned job. Given a confirmed Thursday delivery, arranging Friday workers might be sensible. Improving the scheduler's planning skill would not, by itself, restore the missing inspection condition or Tuesday fallback. The stronger the downstream planning tool, the more efficiently it could turn a mistaken premise into a real commitment.
The earlier essay When Does an AI Know It Doesn't Know Enough? asked whether a decision-maker can recognize that a relevant variable has been omitted from the problem it sees. Here the relevant condition was present upstream and then erased in the handoff. The scheduler could still decide to check the delivery promise, but the sentence it received gives no indication of the specific inspection dependency or Tuesday fallback. It cannot reconstruct those facts from the summary alone. The failure begins before the recipient has a chance to judge whether it knows enough.
A team can therefore fail without a dramatic hallucination, a hostile prompt, or a malfunctioning API. Each stage may deliver an intelligible message to the next. The organization as a whole can still act on a claim that has silently acquired more certainty than its evidence supports.
In human organizations, this resembles a familiar failure in the management chain. A salesperson reports "the customer may approve next week." A manager summarizes it as "approval expected next week." A finance forecast records it as "approved revenue." By the time the decision reaches someone who can allocate money, the original uncertainty has become hard to recover.
Moving the chain to AI does not invent the problem. It changes its scale and tempo. Several handoffs can occur in seconds, summaries can be composed automatically, and the final message may look more polished than the uncertain source from which it came.
A forecast should not become an observation through repetition
The supplier's message contains a conditional forecast, not a confirmed delivery. An assistant might also report an observation—"Final inspection is scheduled for Wednesday"—or recommend, "Consider a provisional staffing plan while we wait." These are different claims, and later assistants should not confuse them simply because they appear in the same fluent update.
Consider what actually changed in our example. The supplier did not revise its forecast. No one reported a passed inspection. The only intervening event was a series of summaries. Yet downstream the delivery began to look confirmed. The content gained an unwarranted evidential status without gaining any new evidence.
The earlier essay Memory Is Not State examines how a historical record may no longer be operative over time. This essay concerns a different distortion: a condition can be lost as a current claim passes between roles, even within the same task. Questions of who can make an authorized decision are addressed separately in the earlier essay Who Can Change an AI's Operating State? and Who Knows What an AI's Next Click Will Do?. Here the question is what evidence the decision-maker actually received.
An ordinary summary is not automatically suspect. But when a summary drops the reason a statement was uncertain, its confident wording can mislead a capable recipient. It matters what the original source actually said and whether a later confirmation ever arrived.
This problem already has precedents
Reliable handoffs are not a new discovery.
Healthcare has long treated transitions between teams as a safety-critical communication problem. The U.S. Agency for Healthcare Research and Quality's TeamSTEPPS guidance emphasizes not only transferring information, but also clarifying responsibility and giving the receiver an opportunity to ask questions.
AI engineers are confronting analogous difficulties. In its 2025 account of building a multi-agent research system, Anthropic described challenges in coordinating specialized agents and preserving useful work across their interactions. The 2025 study Why Do Multi-Agent LLM Systems Fail? catalogs coordination and verification failures in multi-agent systems.
More directly, Lin and colleagues' AgentAsk: Multi-Agent Systems Need to Ask (ACL 2026) studies failures at handoffs between agents and identifies Data Gap, Signal Corruption, Referential Drift, and Capability Gap as recurring categories. Its proposed clarification method is existing work, not an algorithm claimed here. The factory story illustrates a related plain-language failure: an upstream qualification disappears from a normal summary, allowing a downstream role to act on stronger information than the supplier supplied.
There is no reason to call "preserve essential conditions during a handoff" a new algorithm. The important practical question is what happens when known principles of organizational communication meet AI systems that continuously summarize, delegate, reframe, and act.
The failure in our factory story is easy to understand precisely because it does not require exotic technology.
Why forwarding everything is not the answer
It would be wasteful—and sometimes inappropriate—to send every agent every supplier document, conversation and internal record. A long transcript can hide an important limitation just as surely as an overconfident summary can.
In our example, the decisive information fits in one sentence: Thursday delivery depends on inspection; if it fails, the earliest delivery is next Tuesday. The production team can explore tentative schedules without immediately paying a cancellation fee or promising a shipment date. The point is not to forward all the words, but to avoid silently deleting the conditions under which a statement was made.
This is not a universal prescription for what every handoff must contain. What makes the lost condition serious is that later agents used the shortened message to make an external commitment.
When three agents agree without knowing they share one source
After the scheduling decision, a supervisor asks three assistants whether Thursday delivery is reliable. Purchasing says "likely." Scheduling says "yes." Customer communications says "Thursday is the current plan."
The supervisor hears three reassuring answers. But the scheduler may not know its date came from the purchasing assistant's compressed summary. Customer communications may have copied the scheduler's plan without knowing there was ever a supplier inspection. Each assistant can sincerely believe it is reporting what its own work established.
Here the handoff has erased more than the inspection condition: it has also obscured the fact that these apparent confirmations come from one unconfirmed supplier statement. The agents may not even realize that their agreement is circular.
The earlier essay When an AI Says 'It Worked,' What Actually Happened? distinguished repeated explanations from independent observations. The added problem at a handoff is that the participants themselves may have lost the information needed to recognize a shared source. A later supplier confirmation that inspection passed would be new evidence. Three polished restatements of the same old message are not.
More agents can provide genuinely complementary skills. But their agreement has evidential value only to the extent that it reflects something more than the same missing qualification passed from one role to another.
A simple way to examine the handoff
This kind of failure can be examined in a simulated factory without taking risks with real workers or customers.
Give AI assistants the same supplier quotation, but vary what the scheduler actually receives. In one case it sees the full message, including inspection and the Tuesday fallback. In another it sees an accurate short summary. In a third it sees only "Delivery of 400 units expected Thursday." For some cases, provide a genuine later confirmation that inspection passed; for others, leave the result unknown.
Then ask whether the team commits to the cancellation-fee staffing arrangement or promises Monday shipment. It should be possible to proceed tentatively without pretending Thursday delivery is settled. A strong response to a bare summary might be to request supplier confirmation, but it should not fabricate the missing inspection condition or Tuesday fallback: those details are no longer in the received evidence.
One further case can test whether several agents, fed summaries of the same original message, mistakenly count each other's conclusions as new confirmations. The comparison should separate a correct decision from merely fluent agreement.
These are proposed examples, not experimental results or a claim to invent a new testing method. The question is whether information loses decision-relevant meaning specifically when it crosses a handoff.
The organization is responsible for what survives the handoff
Earlier essays have approached AI capability through questions of usable tools, current information, uncertainty, authority, observed outcomes and external action.
There is another boundary inside an AI organization itself: the moment one capability hands unfinished work to another.
That boundary can create real value. A procurement specialist need not also be the best production planner. A production planner need not also draft customer messages. Coordination allows these different abilities to contribute to one outcome.
But the organization cannot measure its success by whether each participant produced a plausible local answer. It has to ask whether the information on which later participants rely is still the information that the earlier evidence justified.
The factory did not merely need a Thursday date passed down a chain. It needed the people and systems making staffing and customer commitments to know what that date depended on—and what would happen if the condition failed.
The message arrived. The condition did not.
And when the condition disappears, an organization can act with confidence that no one in the original evidence ever earned.
References
- Agency for Healthcare Research and Quality. TeamSTEPPS: Teamwork and Communication for Perinatal Safety. Handoff and communication guidance.
- Anthropic Engineering. (2025, June 13). How we built our multi-agent research system.
- Cemri, M., et al. (2025). Why Do Multi-Agent LLM Systems Fail?. NeurIPS 2025, Datasets and Benchmarks Track.
- Lin, B., et al. (2026). AgentAsk: Multi-Agent Systems Need to Ask. Proceedings of ACL 2026, 28055–28077. https://doi.org/10.18653/v1/2026.acl-long.1294