Guide
How to redirect an AI agent while it is running
Steering an AI agent means giving it new direction while a task is running, so its remaining plan changes without throwing away the steps already finished. Good steering shows the team which steps the direction changed, and when two people steer in opposite directions the agent stops and asks instead of trying to satisfy both.
Why stopping and re-prompting is the wrong default
With a single-player assistant the only way to change direction is to stop and write a new prompt. For a task measured in hours that throws away finished work, and it means only the person who wrote the prompt can change it.
Steering is the alternative: a direction is added to the running session, and the agent works out which of its remaining steps still make sense in light of it.
What good mid-run steering looks like
- The direction reaches the agent at its next step, not after the whole task ends.
- The visible plan changes. If the team says "focus on district IT teams instead", the step titled "Profile the head teachers" should be replaced, so everyone can see the agent listened.
- Finished steps are kept. A direction changes what is left, not what already happened.
- The agent says back, in one sentence, what it is changing and who asked, so the room knows what moved.
When two people steer the same agent
The hard case in a shared session is two teammates giving directions that cannot both be followed, such as "focus on schools" and "focus on consumers" within the same minute. An agent that receives both will usually try to satisfy each halfway and produce something neither person wanted, without anyone having made a decision.
The safer behaviour is to hold the session before the next step, show both directions side by side, and let the team choose. Only directions from different people should count as a conflict. One person revising their own direction has simply changed their mind, and their latest word should win.
Directions that should outlive the task
Some directions are one-offs and some are rules: "never quote prices in euros", "always cite the primary source". A rule given to one session is lost when that session ends, and the next session repeats the mistake. Keeping it as a team rule that every later session reads is how a correction stays corrected.
How Vrittara does it
Anyone in a Vrittara session can send a direction while the agent works. Before its next step the agent reworks the remaining plan in light of it, keeps what is finished, and posts one sentence saying what it changed. The plan everyone is watching updates with it.
Before each step the session checks pending directions for conflicts between different people: include against exclude, different focus, length, tone, audience, incompatible numeric targets, or one person overriding another. When it finds one it holds and shows both. The same check runs in the composer, so somebody is warned before they send a direction that clashes. A direction that reads like a standing rule can be kept as a team rule that later sessions follow.
Related guides
Common questions
- Can you change an AI agent's instructions while it is running?
- In a shared agent session, yes. A direction added mid-run reaches the agent at its next step, and the agent revises the steps it has not done yet instead of starting over.
- What happens if two teammates give an AI agent conflicting instructions?
- A well-built shared session holds before the next step, shows both directions side by side, and asks the team to pick one. The agent should not quietly split the difference.
- Is one person changing their mind treated as a conflict?
- It should not be. Only directions from different people conflict. When the same person sends a new direction, their latest one wins.