Ask and Agent: read-only Q&A, then full execution

Ask: understand before you touch

Reach for Ask when the thing you need is knowledge, not a change. It’s read-only by construction - it will trace code, explain a flow, and answer “where does X happen,” but it won’t edit a file or run a command that mutates state. That guarantee is the point: you can point it at code you’re afraid of and know nothing will move.

The tell for Ask is that you can’t yet name the change you want. You’re still asking what is this and where does it live:

ask ›
> where does budgetcli decide a transaction is allowed to post
  against a budget? is the check in this repo, or somewhere shared?

  The budget gate is in src/budget/guard.ts, but the rule it
  enforces - the spend-limit threshold - isn't local. guard.ts calls
  policy.allowsSpend(txn), and that function lives in the
  shared `budget-policy` package, imported by six other services.
  Nothing here changes the threshold; it only reads it.

That answer just changed the shape of the job. What looked like a one-repo edit is actually a change to a shared module a half-dozen services depend on - a high blast-radius change wearing a low-effort disguise. You learned that for free, in a mode that couldn’t have broken anything while you learned it. That’s what Ask is for.

Agent: full execution

Once you can name the change, the read-only guarantee stops being protection and starts being friction. That’s the moment to climb a detent to Agent - the right posture for most work, and the mode you’ll spend the most time in. You describe an outcome; it does the whole loop - reads what it needs, edits across however many files it decides are relevant, runs tests, reads the failures, and fixes them - surfacing risky commands for approval as it goes rather than stopping to plan with you first.

Agent is the correct posture when the change is contained and you’d rather review a finished diff than co-author the approach. A gnarly algorithm that lives entirely in one file you own is a perfect Agent task even though it’s hard - the difficulty is real but the blast radius is small, so letting it run and reviewing the result is cheaper than narrating every step.

agent ›
> the retry backoff in src/rates/fetcher.ts is linear; make it exponential
  with jitter, capped at 30s, and keep the existing max-attempts
  behaviour.

  ⏵ Read   src/rates/fetcher.ts, tests/rates/fetcher.test.ts
  ⏵ Edit   src/rates/fetcher.ts   → exponential backoff + jitter, 30s cap
  ⏵ Run    npx vitest src/rates/
    • ok  0.18s
  Done. Backoff now doubles per attempt with ±20% jitter, capped at
  30s; max-attempts unchanged. Tests pass.

What you’re trusting Agent with is latitude, not correctness - you still review the diff. The judgment is matching that latitude to the stakes, which is exactly where Plan comes in. Next: Plan, clarify and research first.