A stuck agent needs a packet, not a ping.

Stopping is not the same as handing off

An agent reaches a case it should not finish. It notifies someone. The notification says that a case needs help, and sometimes it includes a sentence of apology. The human opens the same systems the agent already touched, reconstructs what happened, and hopes they do not repeat a step that already wrote. The agent did not hand the case over. It abandoned the case in public.

Escalation ladders answer who is called and when. Review queues answer where human judgments wait. Neither one, by itself, says what rides with the case. That missing object is the handback: a packet specific enough that a prepared human can act, and specific enough that the work can return to the agent without starting over.

This note is the TWIMCO Agent Support Handback Spec. It defines the packet for one agent job, not for “AI” in general. It is not a personality guide and not a prompt to be friendlier when failing. Friendliness does not carry state. The spec is also not a live tool on this site. The fields below are the asset. If you are deciding whether a workflow should be agent-shaped at all, the live AI workflow readiness check is a different question, asked earlier.

Spin:Market puts software control next to a physical operation. In that kind of setting, a stuck automated step is not a chat inconvenience. Someone has to know what already moved. The same packet discipline applies anywhere an agent may act and then stop: the human inherits actions, not a vibe. Designing that boundary is AI applications and agents work when the product is the agent, not a wrapper around an unlogged thread.

Write the spec before the agent is allowed to act on real cases. A packet designed after the first bad handoff will be a scar, and scars omit the field that was merely annoying until it was expensive.

What rides with the handback

The TWIMCO Agent Support Handback Spec requires these fields for the job you are specifying. If a field cannot be filled from the systems you have, that is a product gap, not a reason to drop the field.

Case identity. The identifier in the system of record, plus the customer, order, lot, or other handle a human would search. An internal trace id alone is not identity if the human cannot open the record with it.

Why it stopped. A reason from a small list you defined for this job: missing fact, low confidence, permission denied, conflicting sources, tool error, policy hold, or out of scope. Free text is allowed as a note under the reason. Free text instead of a reason cannot be counted, trained on, or staffed for.

Case context the human should not have to rediscover. The facts the agent was using, with sources. Not the whole prompt. The facts. Which document, which row, which message. If a fact was inferred, label it as inferred. A handback that presents inferences as records will be trusted once and then never.

Last actions. What the agent already did, in order, including tool calls that wrote or sent. Attempted actions that failed belong here too. The dangerous omission is a successful write that the packet forgets, because the human will do it again.

Evidence attached. The links, excerpts, or artifact ids a reviewer needs in order to check the stop. Evidence is not a restatement of the agent’s conclusion. If the stop is “sources conflict,” the packet shows the conflict. If the stop is “tool error,” the packet shows the error and whether the tool may have applied the effect anyway.

Suggested next move. One concrete action, marked as a suggestion. The human may reject it. A suggestion that says “please review” is not a move. A move names the step: confirm this field with the customer, choose between these two records, retry this tool, or close as out of scope.

What the agent must not do while waiting. The holds. No further writes, no further customer messages, no silent retry. If something must continue, such as a timer or a lock, name it so the human does not fight a background process.

Owner and age. Which queue or role receives the packet, and how old the case was when it stopped. Age is context for urgency. Do not invent a target time in the spec unless your operation has already chosen one. A blank target is more honest than a slogan.

That packet is the handback. A chat line that says “I need help” is not a filled spec, even if the tone is excellent.

The return path is part of the packet

Handback without return turns the agent into a machine that only creates homework. Some cases should stay with the human until done. Say so. Other cases should return when a named condition is true. The TWIMCO Agent Support Handback Spec includes the return rule on the same page as the packet, because a return invented later will not match what the human was told.

Return names the condition, the role who may send it back, and the state the agent resumes from. Resume from “the start of the transcript” discards the human’s correction. Resume from “the full conversation, unsummarized, with tools free to repeat writes” risks a duplicate effect. The clean version is: here is the decision the human made, here are the fields they set, here is the next step the agent is allowed to take, and here are the steps that stay forbidden.

If the human edits a fact, the return packet says whether that edit is now source of truth or only an instruction. Those are different. An instruction can be ignored by a later retrieval. A source of truth has to be written where the next run will see it.

If the human closes the case, the agent does not get a return, and the spec says who notifies the parties. Silence after a human close is a second kind of stuck.

Partial work by the human is the ordinary case. They answered one question and not the rest. The return rule should allow a partial return: the agent may proceed on the answered part and must not reopen the unanswered part by guessing. If your job cannot be split that way, the spec says the case stays human until all required decisions are present. Do not leave that to the model’s enthusiasm.

The return is logged with the same identity as the handback. Someone later should be able to see the stop, the human decision, and the resume as one story. If those live in different tools with no shared id, the spec is not implemented. It is a document next to a gap.

What the spec refuses

A few patterns show up whenever handback is treated as a notification feature. They are composites from agent delivery, not measurements of any deployment.

The packet is a screenshot of the chat. The human can read the story and still cannot find the record, the write, or the conflict. Context without identity is a novel.

The last actions list omits the tool that succeeded. The human repeats a refund, a message, or a status change. The spec’s last-actions field exists so that omission is a failed packet, not a mystery.

The suggested next move is a paragraph of reasoning. Nobody can do a paragraph. The move should be short enough to accept or reject. Reasoning can sit under evidence. It is not the move.

The return is “the agent will take another look.” That sentence has no condition, no forbidden step, and no place the human’s edit is stored. The case will bounce. Bouncing is not collaboration. It is two parties redoing the same uncertainty.

The queue is whoever sees the ping. The spec’s owner field is a role or a queue with hours and a backup. If you do not know the backup, the packet can be perfect and still wait out the weekend. That is an operations gap the spec should make visible rather than paper over with a louder notification.

Refusal, here, means you do not call the integration done while these are true. You may still pilot on low-risk cases if the packet is filled and a human is actually watching. You may not describe an unlogged ping as a handback.

Nothing in this spec tells you the agent is accurate. It tells you the stop is handlable. Accuracy is a different review. Do not let a tidy packet substitute for it, and do not let a demand for accuracy excuse a packet that makes the human start from zero.

Put this into practice

1. Name one agent job. Write the stop reasons it is allowed to use.
2. Fill the TWIMCO Agent Support Handback Spec: identity, reason, context, last actions, evidence, suggested move, holds, owner.
3. Write the return rule, including partial return and where a human edit is stored.
4. Reject any implementation that can stop without producing the packet. A channel ping may accompany the packet. It may not replace it.
5. If you need the agent built around that boundary, scope it as AI applications and agents. Keep the operational standard concrete, the way Spin:Market has to be concrete when software actions meet a real floor: what already happened must be visible to the person who takes over.

An agent that knows when to stop is only half of the design. The other half is the packet it leaves in the human’s hands, and the rule for what comes back.