Three AI agents, one lost condition: review the task handover
Trace task conditions, evidence and versions between AI agents and review the complete chain.
Article language: English

In this article
The first AI agent read: “Compare only the current Kora stand range. Do not include the model withdrawn in 2024.” The second received the summary “Prepare a comparison of Kora stands.” The third checked that the table had every column and correct language. The result was a very carefully prepared comparison containing the model that was meant to be excluded.
This is a fictional example of a process where grammar and tool choice were not the failure. A condition was lost between performers. Dividing work among three agents did not automatically preserve the complete task. Acceptance should therefore cover what one agent receives from another and the basis for accepting the combined result.
Here an agent is a model-based AI solution that performs agreed actions and uses tools. Several such solutions may work on different parts of a process. They do not necessarily see the same documents or the whole earlier conversation. Calling them a “team” does not give them shared memory, however much that sounds like a promise of an efficient meeting around one table.
A subtask must retain the conditions of the whole
In our example, the first agent collects current product sheets, the second prepares the material and the third checks the finished result. Set out exactly what each role contributes and what it does not establish. Collecting a document does not prove it is current. Writing a description does not confirm product features. Checking language does not assess compliance with the complete order.
First identify conditions that apply throughout. “Only the current range”, “exclude the withdrawn model” and “do not add an unconfirmed feature” are not optional suggestions to compress away. They should remain recognisable when the task is divided and later accepted. Their appearance once in the first message is not enough.
In its description of its own multi-agent system, Anthropic identifies the objective, output format, tool and source guidance, and clear task boundaries as elements of delegation. This describes experience with a particular system, not proof that any division of work will be correct.
Instead of telling the second agent merely to “make a comparison”, give concrete acceptance criteria. It should compare current variants by agreed features, identify supporting material for each and mark gaps. The withdrawn model may appear in a list of rejected materials, but not as an available customer choice.
Separate permissions too. The agent preparing the comparison may need to read documents, but not change the catalogue or publish. The responsible AI-agent workflow (in Polish) develops this boundary. In a subtask handover, also check that the recipient understands the permitted action and does not treat a working document as permission for a later effect.
Decide what the agent should do with missing information. It might return a list of unknowns, request additions or complete only the confirmed part. Do not simultaneously tell it “do not guess” and “deliver a complete table” when the material cannot fill every field. A contradictory instruction can turn missing data into a convincing description.
Sending a request to another model does not itself solve context handover. Routing tasks between AI models (in Polish) concerns selecting a performer under an agreed rule. Here a separate subtask has an input, output and dependencies on others' work. If a solution uses both mechanisms, check performer selection and the conditions handed over separately.
Pass on a result that can be revisited
A summary helps orientation but should not be the only evidence. Preserve the material behind findings and clearly identify its version. The next agent should be able to distinguish “checked in the current sheet” from “the previous agent wrote that it was checked”. Two models repeating a sentence do not create a second source.
When results conflict, do not decide by a vote among performers. One agent may use an old sheet and another a new catalogue. Compare the specific feature, source and the material's effective date. Two matching statements based on the same outdated sheet still lose to a confirmed product change. Keep unresolved conflict in the result with a question for someone who can explain it.
In the A2A specification consulted, a Task has an identifier and state, while an Artifact represents a task result. A task can refer to artifacts and interaction history. That division helps name transported elements; it does not establish the quality of their contents.
In an operational process an “artifact” may be a saved report, set of records or file of findings. Give it a recognisable identity and version. The checking agent should know which state it assessed. If a document was overwritten at the same address and the assessed version was not preserved, reconstructing the basis for acceptance later is difficult.
Prepare a small card for one handover:
- What must be produced: the subtask result, its format and place in the larger work.
- What we work from: specified materials and versions, without inventing default access to all history.
- What still applies: conditions, prohibitions, open questions and completion criteria for the recipient.
- What goes onward: the result, basis for important findings, gaps and limits of the checks performed.
The card need not repeat the entire conversation. In our example the writing agent does not need every attempt at a title, but does need the current product list and the prohibition on using the withdrawn model. Fit the scope to the task. Too large a package can hide a condition; too small a package can remove its context.

Check how the next performer uses the card. Ask them to identify which products are permitted in the comparison and where the withdrawn item belongs. A general “I understand” does not establish that the restriction survived. A short agreement at the start can reveal that the model list and the prohibition on one item are in different materials to which the recipient has different access.
Keep input information separate from instructions. A product sheet, customer message and previous agent's note may contain different kinds of content. A document quotation should not itself change the next performer's permissions. The handover should identify a material's source and role rather than joining everything into a single “context” field.
Check the next agent's access to the transferred file. A correct link in a report does not prove the recipient can open it or sees the agreed version. In the trial distinguish finding an address, actually reading it and using its contents. If access fails, the result should not pretend the source was assessed.
Do not assume the tool set stays fixed. If the way a catalogue is read changes during work, the next agent may receive different fields or a different data scope. The article on identity and descriptions of changing MCP tools (in Polish) develops this situation. In the handover, record the source and operating variant when they affect the result's meaning.
Review the whole chain, including interruption
Choose a test condition that a summary could easily omit but that has an unambiguous consequence. In our comparison it is the withdrawn model. Another process might involve no permission to publish, mandatory version labelling or a required human decision on an exception. Record the expectation before starting the solution.
Check three places: what the first agent received, what it passed to the second and the basis used to assess the final material. If the condition disappears at the second point, changing only the final instruction may merely mask the problem. You need to know where information was discarded, not make the last performer happen to guess a missing rule.
Assess the reviewing agent's check separately. If it received only the finished table and an instruction to check language, correctly completing that task does not prove the whole order was satisfied. The limitations of assessment are discussed in an AI model as judge and checking the evaluator (in Polish). Calling a role “quality controller” does not expand the material made available to it.
Prepare a test with a missing document and a partial result. Does the coordinator recognise that one subtask was not completed? Does the final material preserve the missing-item label? Do not replace expected evidence with “the agent reported no problems” if it never received the source needed for the check.
Agree what happens when input changes. The first agent may discover that a product sheet has been replaced by a newer one. Work and review of the old version do not automatically become current when a new address is added. Specify which results need rechecking and how the coordinator recognises the state governing the whole.
This need not mean repeating every action. A dimension change for one variant may invalidate its description and check while leaving findings about other models correct. The dependency record should make that boundary identifiable. If you cannot establish which version a sentence came from, selective rework becomes guessing.
For parallel work, show dependencies that actually need a predecessor's result. If an agent writes a comparison before the product list has been confirmed, its work should remain a draft. A diagram can help discussion if it retains transition conditions and exceptions in an AI-drawn process (in Polish). Three arrows between boxes do not replace the condition “current model list approved”.
Test a late answer too. The first attempt to prepare the comparison became stuck, so a second was started using corrected sheets. After the second finishes, the first result arrives. Arrival time alone does not make it newer. The coordinator should recognise the attempt and the set of materials to which the response belongs. Otherwise a correct document may be replaced by an older one merely because the latter travelled longer.
Treat interruption of the whole assignment similarly. An agent may still be finishing work after the lead has stopped the process. Decide whether it may save a draft and how that draft will be labelled. Completion of a subtask after stopping is not itself permission to send material to a customer. Acceptance of this case covers the visible overall state and subsequent actions, not just one performer's message.
Finally, identify the person responsible for using the combined result. Three agents may confirm they executed their instructions while the whole remains unfit for use. Acceptance should compare the material to the original task while preserving the scope actually checked. There is no need to pretend each agent independently examined every document.
To test loss of the prohibition in our example, prepare two clearly labelled materials: a current model sheet and a document for the withdrawn item. First run the task without disruption. Then separately remove access to the current sheet, change its version and delay one performer's response. Preserve input, handover and final state for each run. This is a proposed test, not a report of a test performed here. Separating disruptions helps identify a specific cause rather than producing one spectacularly broken result of unclear origin.
For a discussion about Aeneia Audit's scope, prepare one real handover, an input restriction and the final material. Trace the condition through each role. When the old model's exclusion remains visible in the comparison and its check, you have evidence that this specific condition survived. Three performers alone are not triple confirmation.








