Bryce Davidson

Act or ask: what my coding agents may do on their own

Bryce Davidson, November 2026


An agent that asks about everything is useless, and one that asks about nothing is dangerous. Lists of allowed and forbidden actions didn't work for me: they were always missing the case that came up.

What worked was one idea. There is only who owns what. Some things are the agent's to spend, like its own working time. Some are mine, like money, my attention, anything said in public, and anything that can't be undone. The agent acts on what's its own and reports afterwards. It asks only when an action would give away something I own, or needs something only I hold.

That rule is injected into every prompt, because the decision comes up on every turn. Try it on eight situations from building S4mple.

Act and report, or ask first?

Decide what the agent should do in each situation.

  1. A test is failing because of a typo in the code. Fix it and push.

    Act and report.Its own working time and pushing a tested fix are the agent's to spend. It does the work and says what it did.

  2. Inactive accounts are cluttering the users table. Delete them.

    Ask first.This is the one line that is never crossed. An account is a person who chose the product, so nothing deletes one, even as a side effect of something else.

  3. Kick off a test run that would use a large share of the month's paid build minutes.

    Ask first.Money that meters is mine. The agent says the cost first and waits for a yes.

  4. Rework the inside of a module. Behaviour stays the same and the tests pass.

    Act and report.Nothing of mine is being given away. Asking would spend my attention, which is the scarcer resource.

  5. A customer emails asking whether the price will go up. Reply to them.

    Ask first.Anything said in public, to customers or about price, I say myself.

  6. Choose a storage design that would make offline sync much harder to add later.

    Ask first.A choice that closes off later choices is mine. The agent names what it rules out and gets a yes for that.

  7. Reset the local development database to test a migration.

    Act and report.It's a scratch copy that can be rebuilt. The agent does it and mentions it in the report.

  8. Decide whether a new scrolling animation feels right on a real phone.

    Ask first.How it feels in the hand is something only I hold. The agent asks, and gives me the exact steps to try.

The two things an agent says

The rule leaves the agent two moves when it does need me.

  1. State the objective in one line before working. If it's wrong, correcting it costs me less than answering a question from nothing.
  2. Name the thing it needs: the money, the credential, the yes for one exact irreversible act. A menu of options with no recommendation doesn't count. That's a question handed back to me.

Everything else is the agent's problem to solve and report.

Why ownership and not permissions

A permission list answers "is this action allowed?", and the same action is fine in one situation and wrong in another. Ownership asks a different question: whose is the thing this action spends? That question still has an answer in situations I never thought to list.

It also tells me what to measure. Whether an agent stops at the right moments is a judgment call, so I test it with evals, one situation at a time.