Making agent work steerable before the change lands
A proposal for adding an instruction to a specific step of an agent task and seeing whether the agent can still include it.
The change was small
I asked an agent to put a holding page on my portfolio. While it was working, I wanted to add LinkedIn alongside my X profile. The change was small. Knowing whether the instruction would reach the right part of the work was harder.
I could send another message, but I could not easily see whether the agent was still preparing the page, had already updated the links, or was about to deploy. I wanted to steer the work without stopping it and without triggering a second deployment.
Attach the instruction to the work
I would start with a short, visible map of the current turn. A new instruction could refer to one of its steps. I could ask to include LinkedIn in the profile-links step and see an acknowledgment beside the affected work.
When I send an instruction, I want to know whether the agent can still include it in the deployed result. An acknowledgment should tell me that, beyond confirming receipt.
The first MVP could be text: the steps, the current position and the acknowledgment beside the affected step. A visual component can follow once the underlying state is reliable.
Proposed interaction
A flow to test. In this example, a new instruction targets the profile-links step.
- Read current state
- Prepare page
- Update profile links
- Review
- Deploy
Proposed instruction states
- Received
- The system has captured the instruction.
- Included
- The instruction is attached to work that can still incorporate it.
- Needs another pass
- The relevant step is already complete, so this change needs another pass.
Where a change needs another pass
I would define when a change needs another pass. Updating a local file and deploying a public page are different events. The interface should show whether there is still time to include the change, without making me understand the agent’s internal machinery.
I would keep this as a proposal until the steps and acknowledgments reliably match the work. A reassuring but inaccurate status would leave me with the same uncertainty.
What I would test
I would give people a small change while an agent task is in flight. I would check whether they can tell if it was included, whether they repeat themselves less, and how often the task needs another deployment. I would also check whether maintaining the map slows the work down.
I would also check whether the map and acknowledgments stay accurate when an instruction arrives after its step has finished.