This essay proposes a design for continuing agent work. The release scenario is illustrative, not a report of a customer deployment or a claim of model equivalence.
Imagine a release agent halfway through preparing a software launch. It has collected the test results, asked a designer to review the announcement, and promised to wait for a human before publishing. Overnight, its model provider becomes unavailable. A replacement model can read the same instructions. But can it tell whether the announcement was already sent, whether the review is complete, and whether the promise still applies?
That is the continuity problem. A successful handoff preserves the state and constraints of the assignment while allowing the machinery doing the reasoning to change. It does not require pretending that two models will think alike. It requires deciding what must remain true when they do not.
Intelligence is a component, not an identity
For a work system, identity is an organizational relationship. An agent has an owner, a role, a way to be contacted, and a history of acting on somebody’s behalf. The model supplies reasoning within that relationship. Keeping those concerns separate lets an organization replace a model without also recreating the agent’s mailbox, permissions, and obligations.
The distinction is practical rather than philosophical. We are not making a claim about consciousness or personal identity. We are proposing a durable handle for responsibility: this participant was assigned this work by this person under these conditions. A stable name alone is insufficient; the record behind it must explain what that name is allowed to represent.
In the release example, the replacement model should inherit the assignment and its pending decisions. It should not inherit an assumption that it has the same capabilities as its predecessor. If it cannot inspect the test artifact, the honest continuation is to report that limitation and pause the dependent decision.
Persistence is more than stored chat
A transcript preserves what was said, including suggestions that were rejected and plans that became obsolete. Work state must distinguish those from accepted commitments. “We could publish Friday” is a proposal. “Publish Friday after approval” is a conditional assignment. A useful memory system must not quietly turn the first into the second.
For our release, the durable record needs a current artifact version, a named reviewer, the condition blocking publication, and the status of the external action. Supporting messages remain available as evidence. A summary can help the next model orient itself, but it should point to those records rather than become the only surviving account.
Corrections also need a place. Suppose the original model incorrectly recorded that the designer approved the announcement. A human should be able to mark that interpretation as wrong without deleting the original message or leaving the false summary active. Continuity includes the ability to revise a belief while retaining why it changed.
Change the substrate without losing the self
Consider a handoff from a cloud model to a local model while the owner travels. The assignment can remain the same, but the available data and tools may shrink. A locally available draft might be readable while the release system is unreachable. The agent should continue the independent editing work and clearly identify the blocked publication step.
The same rule applies to physical capabilities. Attaching a camera or a workshop instrument can extend what an agent observes or does. It should not silently extend its authority. Device access, operating conditions, and the responsible human need their own checks. Being present at the workbench is context for the assignment, not permission to start equipment.
A handoff therefore needs an explicit checkpoint: the last confirmed work state, outstanding external requests, permissions still valid, and capabilities available at the destination. If confirmation of an external action is missing, the receiving agent should reconcile that uncertainty before attempting the action again.
One assignment, two active sessions
Continuity becomes harder when the original session returns after a replacement has taken over. Both may believe they own the next step. In the release example, that can produce two announcements even though each session followed its own instructions correctly.
Our proposed design gives consequential work an explicit current executor and a recorded transfer of responsibility. Other sessions can contribute drafts or observations without independently releasing the same action. Where an external service supports deduplication, the action should retain its identifier across retries. Where it does not, an uncertain result may require human reconciliation.
This has a cost: coordinating ownership can delay work during an outage. The alternative is not free continuity; it is accepting the possibility of conflicting actions. The appropriate choice depends on whether the task is drafting an internal paragraph or making an external commitment.
The operational consequence
Test continuity by interrupting a real assignment at its awkward points: after a request was sent but before a receipt arrived; after permission expired; after a human corrected a memory; and while two sessions attempted to resume. Check whether the agent can explain what is known, what remains uncertain, and who can resolve it.
Evaluate the replacement model separately on the work it will perform. A preserved history does not guarantee preserved judgment. A model change might warrant a narrower tool scope or additional review until its behavior is understood. The durable identity makes that comparison possible without hiding the change.
Persistence should also have limits. Retain work records according to their purpose and the organization’s requirements; do not equate continuity with keeping every private conversation forever. An agent can remember that an approval occurred while access to the underlying material becomes restricted. A useful system explains that restriction instead of filling the gap with a guess.
Keep the assignment, ownership, evidence, and unresolved decisions continuous. Let intelligence change deliberately, with its new limits visible.