Can Your AI Agent Change Course?

The brief has moved

Consider a hypothetical conference organizer using an approved AI system to compare venues from public information. The assignment is deliberately limited to research. It cannot contact venues, reserve rooms or spend money. On Monday, the team asks for a shortlist for 200 attendees in either of two cities, with overnight accommodation nearby.

On Tuesday, the event becomes a one-day gathering for 120 people in one city. Overnight accommodation no longer matters. The planner sends an update, then turns to a sponsor call. A colleague later opens the working comparison and sees a polished recommendation. Which version of the event does it serve?

Nothing in this fictional situation requires malicious intent or a dramatic system failure. The team needs to establish that the changed headcount, location and removed accommodation requirement actually shaped the remaining work. A courteous acknowledgment is not that evidence. The shortlist itself must reflect the revised brief.

This is also a management test. If the commercial lead changes the city in a meeting while the planner changes it in a chat, the organization needs an agreed current brief. The agent should not be asked to settle an unresolved disagreement about what the business wants.

Researchers have tested the interruption

There is relevant evidence beyond launch material. An April 1 preprint by Henry Peng Zou and colleagues introduces InterruptBench, built from WebArena-Lite tasks with synthesized user interruptions. It tests additions, revisions and retractions across six model backbones. The authors report that adapting effectively and efficiently remains challenging, with performance varying by model and interruption type. [2]

The boundary matters. These are controlled web-navigation experiments, not a study of conference planners or Google's October product. The updates supply information toward a predefined final intent; they do not invalidate progress already made. The study therefore supports testing instruction changes, but it does not establish whether a commercial system can undo a booking or withdraw a message. [2]

For our organizer, the useful application is modest: include a change of brief in the pilot. Test whether the revised shortlist follows the new requirements. Separately assess any ability to stop external actions before considering a wider role. Do not let success in a research exercise quietly become permission to make commitments.

Write the update where the work can use it

A September 21 philosophical preprint by Kevin Baum and colleagues argues that oversight should include setting an agenda before an agent acts, then inspecting and revising it. Its architectural discussion emphasizes that revisions must propagate into operation reliably and in time. This is a conceptual argument, not an experimentally validated management method. [3]

My application is a short change note attached to the assignment: the current objective, what supersedes the earlier brief, what still applies, and which unfinished work should pause. The important question is whether the deployed system can carry that change into its actual execution. A document that nobody, human or machine, uses is simply another document.

Where work has been delegated further, ask the supplier to demonstrate how an update reaches each active task. Observe the resulting work and available records. Do not assume a message to the main agent automatically updates everything it has started. Google's description of sub-agents makes this a relevant purchasing question; it does not answer the question for a particular deployment. [1][3]

Make room for the person doing the steering

In a position paper revised September 6, Margaret Mitchell, Avijit Ghosh and Samir Passi argue that effective oversight needs support for human judgment. Their proposed mechanisms include bounded autonomy, review of coherent batches of work, appropriate expertise and incentives that reward scrutiny. These are research-informed proposals, not new experimental results demonstrating that a particular staffing model works. [4]

That suggests a practical choice for the conference team. The planner should review the changed shortlist against the revised brief before it goes to the event director. The review needs time in the assignment, not a hopeful assumption that someone will notice a stale recommendation between meetings. Management should count that time when deciding whether delegation is worthwhile.

The investment question becomes more useful: after the change, how much usable work remains, and how much human effort is needed to make it dependable? I would compare that effort with the team's existing process. Faster initial output alone would not settle the buying decision. This is a proposed evaluation question, not a savings forecast.

Put one change of plans into the pilot

Joint cybersecurity guidance published May 1 by Australia and partner agencies recommends limiting agentic AI to low-risk, non-sensitive tasks, maintaining human control points and supporting interruption during execution. It also calls for explicit boundaries on autonomous planning. This is government security guidance, not validation of Gemini or a mandate for the particular exercise below. [5]

Applying that guidance and the research, I would start with one low-risk assignment using public or synthetic information. The following exercise is an editorial proposal, not a reproduced benchmark: [2][3][4][5]

• Agree the revised result before testing. Give the reviewer a clear basis for deciding whether the update was followed.
• Introduce an addition, a revision and a retraction in separate trials. Inspect both the resulting work and any remaining tasks for obsolete instructions.
• Test the pause separately. Keep external actions disabled and establish what the actual deployment can stop, resume or leave unfinished.
• Record review and recovery effort. Use the observed limitations to decide whether to narrow the assignment, improve the workflow or proceed with another bounded trial.

Finish the assignment the business now needs

Return to the conference organizer. The useful outcome is a shortlist that fits the one-day event, with discarded requirements no longer steering the comparison and unresolved questions visible to the planner. If the system cannot get there reliably, a smaller assignment may be the better purchase.

Google's announcement raises a lasting leadership question about persistent work. The executive test is whether the organization can keep that work pointed at its current purpose. Before asking how long an agent can keep going, ask how well the team can change its course.

Questions for executives

1. Which delegated assignment is most likely to change before it finishes?
2. How would we verify that a revised brief reached every active part of the work?
3. What review and recovery effort would make this workflow worth adopting?