The situation
Asking too little and asking too much both cost you.
An agent that never asks makes calls that aren't its to make. It picks a database design at the end of a long task, adds a library nobody reviewed, or reads data it had no reason to touch. The work looks finished. The decision inside it was never anyone's.
An agent that always asks has the opposite problem. Every question pulls a person out of their own work, and after a week of small questions people stop reading them. That is a slower way of doing the work yourself.
The right line isn't a feeling each agent guesses at, run by run. It's a decision you make once, for each kind of decision.
The line
Three kinds of decision, and who makes each.
Most decisions an agent meets fall into one of three kinds. The question is never whether agents should ask. It's which kind a decision belongs to.
The agent decides
Changes that stay inside the task, can be undone, and are covered by tests.
- Rename or refactor inside a module
- Add or change tests
- Fix a failing build
The agent proposes, a person approves
Choices other people will build on. The agent lays out the options; a person picks one.
- Change a database schema
- Change a public API
- Add a new dependency
A person decides
The agent stops before it starts, and waits for someone accountable.
- Touch payment or billing code
- Read or move customer data
- Change who can sign in, or what they can do
Why it drifts
Four signs your line is in the wrong place.
A line drawn once still needs checking. These four patterns show up in reviews long before anyone calls them a problem.
The same question comes up every cycle
If agents ask it every time and the answer never changes, it's a rule, not a decision. Write it down once and stop asking.
Make it a ruleReviews keep reversing an agent's choices
When people undo the same kind of call again and again, the agent should have asked before it made it.
Move it towards a personPeople approve without reading
With too many asks, approval becomes a click. The line has drifted too far towards the person.
Let the agent decide moreNobody can say who approved a change
An answer given in a chat thread isn't a record. If the ask wasn't recorded, the decision can't be traced.
Record every askThe questions
Five questions for your next review.
Put these to your tech leads. If the answers differ from team to team, your line is being drawn by whichever agent runs next.
Which decisions can an agent make today without anyone knowing?
VisibilityIf you can't list them, you haven't drawn a line. The agents have, one run at a time.
What does an agent do when the ticket is ambiguous?
AmbiguityGuessing produces work that looks finished and isn't. Asking only helps if the question reaches someone who can answer it.
Who answers an agent's question at 7pm?
OwnershipAn ask with nobody to answer it is a stalled piece of work. Every kind of ask needs an owner.
How many of last month's asks changed the outcome?
SignalAsks that never change anything are noise. Reviews that keep changing things mean the agent should have stopped sooner.
Where is each answer recorded?
RecordIf the answer lives in a chat message, the next agent will never find it, and neither will the next auditor.
Try this on your organisation
Draw your line.
For each kind of change, choose who decides. The page turns your choices into a stop-and-ask policy you can paste into your team's working agreement. Nothing you choose leaves this page.
There's no single right answer. A line your teams agree on, written down, beats a better one each agent guesses at.
With Wakalix
An ask is part of the run, not an interruption.
In Wakalix, where an agent stops isn't left to the agent's judgement in the moment. It's part of how the agent is defined, and every answer is kept.
- 1A plan runs only after a person approves itNothing runs until someone on your side approves the plan. Approving it is what creates the work.
- 2It stops, lays out the options and waitsWhen an agent reaches a decision that belongs to a person, it stops, lays out the options and waits. It doesn't guess. Stopping to ask is a defined outcome of the agent's run, and its checkpoints mark where a person decides.
- 3A person accepts the workEvery piece of work comes back with its evidence. A person accepts it, challenges it or sends it back, and every agent action is on the audit trail.
- 4The retro moves the lineWhat an agent refused, got wrong or had to ask about goes back into the agents themselves, so the next cycle starts from a better team.
Your engineers keep the agent harness they already use. Wakalix never runs the model.