The Work Must Outlive the Session
Imagine the model session in front of you disappears tonight.
You come back to the work tomorrow. The transcript may still open. A draft may still be sitting where you left it. But is it the current draft? Was the last confident paragraph an accepted decision, a suggestion or an assumption nobody had challenged yet? Which source was still in dispute? What had the work been permitted to become?
This is the point where a smooth conversation reveals its fragility. The words survived. The status of the work did not.
It is tempting to call that a memory problem. We think that gives the session too much credit and the production too little responsibility. The model did not forget the work. The work was placed in the wrong owner.
A capable session can reason, write, inspect and revise. In our experience, its active context can hold texture that a folder of finished files does not record. But the session is a working context. It is not the institution around the work.
Our operational conclusion is that long-horizon work becomes manageable when authority and continuity belong to the production system rather than to any individual model session. Manageable does not mean certain to succeed. It means that the work can be recovered faithfully enough to judge, stopped when permission runs out and continued without pretending that momentum is authority.
The morning after the conversation
We arrived at this distinction through failure, not vocabulary.
In our bounded experience, a mixed-provider production was deliberately interrupted while work was active. An earlier interruption attempt had orphaned workers. We kept the evidence of that failure and proved the repair in a separate controlled run instead of tidying the first attempt into a success story. In the repaired run, the work stopped cleanly, its production state remained recoverable and the same authoring contexts continued. The finished result still had to pass independent verification. Nothing about successful recovery made it public.
The important part is not the recovery mechanism. It is what the production refused to lose.
It retained the commission that defined the work, the material already produced, the failure that still counted and the authoring contexts that could continue without posing as fresh substitutes. It also retained a negative fact: completion was not publication permission.
We have seen the same boundary in quieter moments. In our bounded experience, when the production reached a question only we could answer, consequential work stopped until our answer could return to the context that raised it. The work did not fill the silence with a plausible substitute. It waited because the missing thing was not information in general. It was authority.
That is the human problem hiding inside session continuity. On the morning after a conversation, we do not merely need to know what the model said. We need to know what remains true, what remains open and what we are allowed to do next.
What the production has to carry
We use separate working definitions for durable material state, resumable authorship and authority because they solve different failures.
Durable material state is what exists outside an invocation: the governing commission, the current artifact, the admitted evidence, the decisions already made, the failures still open and the status of the work. If the session closes, these things do not become guesses.
Resumable authorship is different. It is a truthful route back to the authoring context that produced the work when revision depends on the reasoning held there. In our experience, a file can be perfectly preserved while the live argument behind it is flattened into output. A replacement can edit the sentences. That does not make the replacement the same authoring context.
Authority is different again. It is who may define the work and who may change its status. A person or system can possess the artifact without permission to accept it. An authoring context can resume without gaining the right to approve its own work. A reviewer can hold authority and still be unable to exercise it responsibly if the evidence or exact candidate is missing.
In our interpretation, possessing any one of these does not confer the others. A saved artifact has no permission. A resumed author has no automatic vote. An approval detached from the object reviewed is only a reassuring label.
This is why continuity is not one synthetic mind remembering everything. Our working definition is faithful recoverability: which commission governs, which artifact is current, what evidence supports it, what failed, what was decided, which authoring context should continue and which questions still belong to the operator. Different contributors do not have to become the same intelligence. They have to remain answerable to the same production.
Variety is not quality. Answerability is the point.
Authority lives in the verbs
It is easy to imagine authority as a final approval box. By then, many of the consequential decisions have already passed.
An assumption becomes a requirement when the authorised commission adopts it, not when a session repeats it confidently. An output becomes evidence when its origin, relevance and limits can be inspected against the claim it is meant to support. A draft becomes the selected draft through a real selection decision, not because its filename looks final. A verified candidate becomes an accepted artifact only when the named authority accepts the exact candidate reviewed. An accepted artifact becomes public through another explicit permission.
Our interpretation is that authority lives in those verbs. Each one changes the status of the work. None is supplied by fluency.
Systems can propose, execute, criticise, compare and verify. A validator can establish that a specified condition passed. It cannot invent the authority that made that condition decisive. We have seen technically successful and independently audited packages remain unpublished because publication was a separate decision. In our bounded practice, approval has been bound to the exact material reviewed; a promising name or status label proved nothing on its own.
This does not make the human at the gate infallible. Our interpretation is that a person has to be able to inspect the relevant evidence and the exact object, understand enough to intervene and remain accountable for the decision. Otherwise human review becomes theatre: a signature added to a conclusion the reviewer was never equipped to challenge.
So when we say authority belongs to the production system, we do not mean that software becomes the moral or editorial authority. We mean the production preserves and enforces the route by which human authority acts. The permissions of the work cannot be impersonated by whichever session happens to be running.
Evidence gives the work a way back
The forward path is easy to admire: commission, action, artifact, review. Things keep appearing, so the work feels productive.
The return path is less glamorous and more revealing. When review finds an unsupported claim, the finding needs to travel back to that claim and the obligation it failed. When an artifact is missing, the work needs to return to the requirement that called for it. Our interpretation is that evidence gives revision somewhere real to begin. It does not guarantee a good revision. It prevents critique from evaporating into a vague request to do better.
Authority still decides what happens next. The work may be revised, rejected, recommissioned or stopped. Continuity keeps the finding, the failed obligation and the relevant authoring context recoverable long enough for that decision to matter.
This is also why failure belongs in the production record. In our bounded experience, stale status language survived in one of our own design records, and a durable local index once disagreed with the live inventory. Persistence did not resolve either discrepancy. It merely preserved it. Our conclusion is that continuity should keep failure and disagreement visible, not smooth them into a story of uninterrupted progress.
The technical lesson is smaller than the vocabulary
The external sources linked here and in the limits below were accessed in September 2026.
Engineering practice supports this argument, but it does not hand us a universal model for it. Established workflow systems specify materially different recovery contracts: some replay recorded history, some retry or resume checkpointed tasks, and some restore stream state against replayable inputs. Each mechanism has conditions, and none restores meaning, authorship or permission merely by restoring computation. The useful lesson is not that editorial work should copy one of these systems. It is that continuity has to be designed around the kind of state being recovered. Temporal documentation, Airflow documentation, Flink documentation
Provenance work makes a related distinction. The W3C model separates the thing produced, the activity that produced it and the agent bearing responsibility. A provenance record can help someone judge quality or trustworthiness; it does not perform the judgment. That is a useful technical analogue for a production in which artifacts, authorship and authority must remain related without being collapsed. W3C PROV-DM
Current agent research establishes the reliability problem more clearly than it establishes our solution. Performance falls as the human length of a task grows, and messier environments remain harder even at similar lengths. The evidence we reviewed does not show that adding durable artifacts and a human gate is sufficient to produce reliably successful long-horizon work. Our thesis remains an operational interpretation, not a benchmark result. Measuring AI Ability to Complete Long Software Tasks
That is enough literature for the claim we are making. The work needs somewhere more durable than the conversation to live. Research does not absolve us from deciding what that place may do.
Where this argument stops
Our interpretation is that small, disposable tasks do not always justify production-system overhead. Sometimes a short conversation and a disposable result are the honest shape of the work. There is no universal threshold we can give you.
For consequential work, more machinery still does not create quality. A production can preserve a bad premise perfectly. More stored state does not criticise what it stores. More agents do not necessarily repair a weak assignment; studies of coding, simulated workplace and laboratory reasoning tasks have found failures from split information, lost progress and shallow verification, with agent interaction degrading outcomes in some tested settings. Those studies do not test our article production, but they do defeat the idea that added contributors are proof of better work. Why Do Multi-Agent LLM Systems Fail?, TheAgentCompany, HiddenBench
Human gates carry their own failure mode. In flight-simulation experiments, people sometimes followed an imperfect automated aid despite valid contrary indicators; a follow-up found less automation bias when participants were made accountable. This is analogical evidence, not an editorial trial, and we found no primary editorial study showing the same effect in AI-assisted article review. Our interpretation is narrower: showing a person a machine verdict is not the same as equipping that person to judge the work. Does automation bias decision-making?, Accountability and automation bias
A production system can preserve the conditions for judgment across time. It cannot replace judgment. A durable session is useful; it is not an authority boundary. A complete record is useful; it is not truth. A passed check is useful; it is not permission.
Take the work in front of you and apply one test. If its current model session disappeared now, which facts, artifacts, decisions, failures and permissions could you still recover—and who would be authorised to move the work forward? The answer tells you whether the work belongs to a production, or merely to a conversation.