# Agent task scope card

Version: 1.0
Updated: 2026-10-02
Method: https://suhailmohebi.com/notes/how-to-scope-an-ai-agent-task/

Purpose: turn a loose request into a bounded task before an agent acts, so it can move fast inside the edge and pause at it.

Output: one runnable task brief with an observable outcome, a two-way boundary, explicit authority, chosen evidence and stop conditions.

## Use when

A request such as “fix the page”, “research this” or “make it better and ship it” is about to reach an agent that can read files, call tools or change external state.

## Quick start

1. Write the outcome as a state someone can observe, not as an activity.
2. Answer the five parts in order: outcome, boundary, authority, evidence, stop.
3. Size the brief to the risk, paste the copy-ready brief, and park follow-ups as questions.

Prompt polish cannot repair a missing decision boundary. Settle the five parts before improving the wording.

## 1. Outcome

> When the task is complete, [reader, operator or test] can observe [state] that is not true now.

- Main outcome (one):
- Supporting checks:
- What an honest partial result looks like:

Activity: “Review the navigation.”

Outcome: “A first-time reader can reach the guide library from the header, and every destination resolves.”

Several competing outcomes usually mean several tasks. Split them before you brief.

## 2. Boundary in both directions

| Part | Entry |
|---|---|
| In scope: surfaces, records or decisions this task owns |  |
| Must stay unchanged: tempting adjacent work |  |
| Sources of truth: documents, manifests, live state or rules that outrank assumptions |  |

Two correct edits to two different copies can still produce the wrong result. If the authoritative source cannot be identified cheaply, make orientation the first subtask and return after it. Use the live-truth curl checklist for that pass, and the verification checklist when the claim or its proof is still vague.

## 3. Authority, separate from scope

Scope says what the task concerns. Authority says which actions the agent may take. An agent may inspect a whole system and still be allowed to edit only one file.

| Action type | May | Must ask before | Must not |
|---|---|---|---|
| Change: files, records, configuration |  |  |  |
| Decide: design, wording, trade-offs |  |  |  |
| Publish or send: deploy, post, email, message |  |  |  |
| Spend or delete: money, quota, data, branches |  |  |  |

Anything not listed defaults to “must ask before”. “Must not” always includes exposing private material, inventing facts, broadening the request and bypassing a failed safety check.

## 4. Evidence chosen before execution

Decide the claim and its proof together. Proof chosen after the work tends to favour whatever the agent already produced.

| Claim: what must be true | Evidence a reviewer would accept | Lightest reliable method | What it does not prove |
|---|---|---|---|
|  |  |  |  |

- Content change: the rendered page and working destinations.
- Data change: a count and a sample.
- Behaviour change: a real interaction, not a source-code assertion.

## 5. Stop conditions

- [ ] Two sources of truth conflict.
- [ ] The only fix would expand beyond the named scope.
- [ ] The next step is a “must ask before” action.
- [ ] [N] attempts at the same step have failed.
- [ ] Task-specific condition:

On stop, return the evidence gathered, the blocker and the smallest decision needed next. A good stop keeps the progress and leaves a clean decision point for a person or another agent.

## 6. Brief weight matched to risk

| Task shape | Brief weight | Emphasis | Typical stop |
|---|---|---|---|
| Read-only | Light | question, sources, output format | evidence incomplete or sources conflict |
| Reversible edit | Standard | exact surface, acceptance claim, related checks | the change needs broader ownership |
| External action | Heavy | authority gate, recipient or destination, final preview | pause immediately before the action |
| High-impact change | Heavy | risk owner, rollback, independent evidence | any safety or approval condition fails |

Precision rises as reversibility falls and impact rises. Money, personal data, production state and external communication always take a Heavy brief. Length is optional; explicit decisions are not.

## 7. Copy-ready brief

```text
OUTCOME: When done, [person or test] can observe [state that is not true now].

IN SCOPE: [surfaces, records, decisions this task owns]
UNCHANGED: [adjacent work that must not move]
SOURCES OF TRUTH: [documents, manifests, live state, rules]

MAY: [read, compare, edit named files, build, test, draft]
ASK BEFORE: [deleting, spending, publishing, messaging, changing external state]; any action not listed here
MUST NOT: expose private material, invent facts, broaden the request, bypass a failed check, [task-specific]

CLAIM: [what must be true]
EVIDENCE: [what a reviewer would accept], produced by [lightest reliable method]
PARTIAL: [what an honest partial result looks like]
NOT PROVEN BY THIS: [what the evidence does not show]

STOP IF: sources conflict; the fix needs wider scope; the next step needs approval; [N] attempts fail; [task-specific]
ON STOP: return the evidence gathered, the blocker and the smallest decision needed.

NEXT QUESTIONS: record adjacent work you notice as a question with its evidence. Do not start it.
```

## 8. Keep the next question, not another task

When the agent notices adjacent work, it records a question instead of widening the task.

| Question | Why it matters | Evidence so far | Who decides |
|---|---|---|---|
|  |  |  |  |

If the same boundary keeps working across tasks, ask whether it should become a durable project rule, for example in the repository’s AGENTS.md, rather than being copied into every brief.

## Worked example

Loose request: “Improve the documentation landing page and publish it.” Publication is an external action, so the brief is Heavy and the gate sits immediately before production.

```text
OUTCOME: When done, a first-time visitor can choose between setup, reference and troubleshooting from the landing page.

IN SCOPE: the landing page and its data record
UNCHANGED: article URLs and the global design system
SOURCES OF TRUTH: the navigation manifest and the live landing page

MAY: edit and test locally, prepare a draft deploy
ASK BEFORE: production publication to [the docs site’s production URL]; hand over the draft URL and screenshots as the final preview
MUST NOT: expose private material, invent facts, broaden the request, bypass a failed check, rename or move articles

CLAIM: all three routes render and every new destination resolves
EVIDENCE: the draft checked at mobile and desktop widths, plus a link check of every new destination
PARTIAL: the routes render but a destination fails; report which
NOT PROVEN BY THIS: that first-time visitors choose the right route

STOP IF: the navigation manifest and the live landing page disagree; the three routes need changes to other pages; the next step is production publication; 2 attempts fail
ON STOP: return the evidence gathered, the blocker and the smallest decision needed.

NEXT QUESTIONS: two reference articles have no troubleshooting counterpart. Should they get one? (recorded, not started)
```

Verdict: runnable. The agent may finish the draft on its own; production publication waits for a person. The brief does not prescribe implementation steps, so judgement stays with the agent while completion, drift and escalation stay visible.

## Stop if

- The outcome cannot be written as an observable state, or the requester has not ranked competing outcomes.
- No source of truth can be identified, or two sources disagree and nobody owns the conflict.
- The task needs authority the requester cannot grant, or an action nobody has classified.
- The claim cannot be proven with evidence the agent can honestly produce.