Mid-Task Agent Intervention Protocol
Added October 2026
Give human overseers a way to redirect, pause, or correct an agent while it is mid-task, without having to invoke a full kill switch or wait for the task to finish.
Objective
Close the gap between letting an agent run unsupervised and stopping it completely, so a human who notices an agent drifting off course can correct it in place instead of choosing between doing nothing and shutting it down.
Maturity Levels
Initial
The only intervention available once an agent starts a task is to let it finish or kill it entirely. There is no way to redirect it mid-task.
Developing
Some agents can be paused by an engineer with direct system access, but there is no standard interface a non-engineer overseer can use, and pausing does not let them steer the agent back on course.
Defined
A documented interface lets a named overseer pause, redirect, or inject a correction into a running agent task, visible in the same place they monitor the agent's actions. Response time from flag to intervention is tracked.
Managed
Intervention actions are logged and reviewed alongside kill-switch events. Patterns in when and why overseers intervene feed back into agent instructions and permission scope to reduce repeat corrections.
Optimizing
Intervention points are built into the agent's task design up front, so natural checkpoints exist for review before high-impact steps. Intervention data feeds autonomy-expansion decisions under AGT-017.
Get the free AI Governance Control Tracker
Get the free Excel tracker for all 132 governance controls. Score your maturity on Mid-Task Agent Intervention Protocol and every other control, assign owners, and set deadlines.
- 132 controls in Excel
- Score maturity and assign owners
- Track deadlines and regulation coverage
Includes AI Governance Weekly every Thursday. Unsubscribe anytime.
Evidence Requirements
What an auditor or assessor would expect to see for this control.
- —Documented intervention interface showing pause, redirect, and rollback-and-resume actions and where overseers access them
- —Named overseer role assignments for each agent or agent category covered by this control
- —Response-time records from flagged concern to actual intervention, against a stated target for high-impact task categories
- —Intervention logs showing who intervened, why, and the agent's state at the time, for the past 12 months
- —Review records showing intervention patterns were fed back into agent instructions, permission scope, or autonomy-expansion decisions
Implementation Notes
The gap between watching and stopping
A binary kill switch (AGT-012) answers "should this agent be running at all?" A pre-action approval gate (AGT-005) answers "should this specific irreversible step happen?" Neither one helps an overseer who is watching an agent in real time, sees it starting to go the wrong way on a task that is otherwise fine, and wants to say "stop, do it this way instead" without ending the whole task. Several 2026 incidents followed this exact pattern: an agent correctly identified its goal but began executing it in an unintended or unsafe way, and the only available responses were to let it continue or terminate it outright, discarding any useful progress already made.
Where this fits next to AGT-029
AGT-029's go-live review decides whether an agent should start a task or be deployed at all. This control is about what happens once an agent is already running and something needs to change mid-flight, a different concern at a different point in the agent's lifecycle.
What the protocol needs
- A visible intervention point. Wherever overseers already watch agent behavior (AGT-006's audit trail, AGT-011's monitoring dashboard), put pause and redirect controls in the same place, not in a separate tool that requires switching context.
- Three intervention types, at minimum:
- Pause: hold the agent's next action without ending the task, so an overseer has time to think.
- Redirect: inject a correction or new instruction into the agent's context for the remainder of the task.
- Rollback-and-resume: undo the agent's most recent action (where the action is reversible) and let it continue with revised guidance.
- A named overseer role, not "whoever is on call," so there is no ambiguity about who can intervene and who is accountable for having done so, consistent with AGT-027's ownership register.
- Response-time tracking. Measure the time between a flagged concern and an actual intervention, and set a target for high-impact task categories.
- Logging as a first-class citizen. Every pause, redirect, and rollback is logged with who did it, why, and what the agent was doing at the time; this feeds both incident review and the autonomy-expansion criteria in AGT-017.
Keeping it lightweight
The whole point of this control is to sit between "do nothing" and "kill it." If invoking it is as heavy as a formal escalation, overseers will default back to one of those two extremes. Favor a one-click pause and a simple text-based redirect over a multi-step approval workflow for the intervention itself.
Example Implementation
Logistics company whose customer-service agent handles refund requests and occasionally begins approving refunds outside policy
Mid-Task Intervention Log (extract)
| Timestamp | Agent task | Intervention | Overseer | Response time |
|---|---|---|---|---|
| 2026-09-18 10:42 | Processing refund for order #88213 | Redirect: "Confirm return received before approving" | R. Osei, CS Ops Lead | 40 seconds |
| 2026-09-22 15:03 | Batch refund review, 40 orders | Pause: flagged unusual approval pattern on 6 orders | R. Osei | 1 min 10 sec |
| 2026-09-22 15:12 | Same batch, resumed after review | Rollback: reversed 2 of the 6 flagged approvals | R. Osei | n/a |
Outcome: the pattern (approving refunds for orders flagged "return in transit" rather than "return received") was traced to an ambiguous instruction. The instruction was corrected and no repeat occurred in the following 30 days; this incident was cited in the next AGT-017 autonomy review as a reason to hold the agent's refund-approval ceiling steady rather than raise it.