Can a Human Finish the Work When the Agent Stops?
$ grep -n "^##" 2026-11-agent-workflow-manual-fallback.md
A manual fallback is usable only when the operator can establish what happened and take over without competing with the automation.
“A human can finish it” leaves some awkward questions unanswered. Where is the approved input? Which actions have already taken effect? Can the agent resume while the operator is working? Does the operator have access to a supported interface for the remaining action?
I would make a successful takeover drill a condition of delegating a workflow with consequential writes. A button labelled “Escalate” is not enough evidence that the fallback works.
Start from the interrupted operation
Consider a synthetic supplier-onboarding workflow. The agent has saved approved account details, submitted an account-creation request and lost the response. It has not attempted the invitation email.
The operator should see a view of the existing workflow record:
| Item | State at handoff |
|---|---|
| Approved input | Saved account details, with their version and approval reference |
| Account creation | Outcome unknown; original request identity and target provider recorded |
| Invitation | Not attempted |
| Execution owner | Takeover pending; manual writes are not yet enabled |
| Next safe action | Inspect the provider's operation status or reconcile the account request |
| Remaining authority | Complete this approved account and invitation; no other account changes |
These are illustrative fields, not a completed incident report. The references must open records the operator can access without asking the failed agent to reconstruct them.
Transfer control, then resolve uncertainty
First prevent the old executor from submitting competing writes. The mechanism must work where writes are issued: a supported pause-and-drain operation, revoked execution access, or an ownership check enforced by the write path. Choose what the system supports. Updating an owner label in a document does not enforce it.
Then resolve work already in flight. Stopping the local executor does not retract a request the provider accepted. AWS documents a concrete instance of this limit: when a Step Functions .sync task is aborted, cancelling the integrated task is best effort, and the underlying work can continue.
For the supplier example, look for the account associated with the original request. A matching name alone is weak evidence, and an empty search may be inconclusive. A documented, still-valid idempotency contract may permit replaying the original request safely; otherwise preserve the unknown state and escalate it. The retry decision survives the move from agent to human.
Only after confirming that the approved account exists should the operator send the authorized invitation. If ownership cannot be enforced or the remote outcome remains unresolved, leave the affected action blocked with a named owner. Do not make the status look tidier by calling it failed.
Google's incident-management guidance calls for an explicit, acknowledged handoff and clear responsibility for operational changes. Applying that discipline to agent recovery is a design choice: the incoming operator accepts responsibility, and everyone can see who may act next. It does not require building an incident-command system around every task.
Rehearse the fallback before it matters
Use test data to pause a workflow after one confirmed effect and simulate a lost response to another. Give an authorized colleague the handoff and the supported UI or API. Have them identify what to preserve, what to investigate and what they may execute next.
Record manual results, including returned provider identifiers, in the same workflow record before re-enabling automation. A private rescue spreadsheet creates another version of what happened. An invitation failure must not cause an already-created account to be created again.
The drill passes when the operator can complete the remaining authorized work without duplicate effects, or reach a justified stop with the unresolved operation and next owner explicit. If they need the agent's conversation to guess which action happened, the fallback still needs work.
$ subscribe --newsletter
Practical AI engineering, in your inbox
Field notes for technical leaders building agents, evaluation systems, governance, and production infrastructure.
Related
What Belongs in Agent Memory
A durable handoff separates observations, hypotheses and unfinished work, then records what would make each observation stale.
When Should an Agent Ask Permission?
A useful agent acts within delegated authority, prepares consequential actions for review, and asks again when the action changes.
The 30 Principles for Agentic Engineering — Part 2: The Lifecycle
Principles 6–14. How work moves through an agentic engineering team: the ticket as contract, AI distillation with human curation, three gates, verification before done, characterisation tests, the 1.2× capacity rule, the J-curve, and telemetry.