Research Note

AI Coding Task Autonomy Matrix

Task suitability depends on evidence and consequence, not on whether an agent can generate a plausible diff.

Aug 4, 20262 min readBy Dalton Anderson
In this article

AI Coding Task Autonomy Matrix

Decision principle

Task suitability depends on evidence and consequence, not on whether an agent can generate a plausible diff.

Dimensions

DimensionLower-risk signalHigher-risk signal
ClaritySpecific outcome and examplesAmbiguous product or policy decision
ReversibilityIsolated diff and clean rollbackData deletion, migration, or external mutation
ObservabilityFast deterministic testsDelayed, noisy, or unavailable feedback
TestabilityKnown acceptance checksCorrectness depends on hidden real-world state
AuthorityLocal read and bounded writeSecrets, production, users, money, infrastructure
SensitivityPublic or synthetic dataPersonal, regulated, confidential, security data
Blast radiusOne module or sandboxShared platform, identity, billing, security boundary
ExpertiseReviewer understands the domainNo qualified reviewer can validate the result

Autonomy levels

Research or explain is read-only and produces no repository mutation.

Propose-only may create a plan or patch for review but cannot execute commands or apply changes.

Supervised execution allows bounded local edits and commands with visible checkpoints and immediate review.

Delegated execution is limited to tasks with strong isolation, permissions, tests, logs, and rollback.

Hold means the task requires a qualified human decision, stronger controls, or new authority before agent work can begin.

Scope-change rule

Reclassify the task whenever the agent discovers a migration, new dependency, external service, secret, production system, payment, user data, authorization decision, or broader shared interface.

An originally low-risk documentation change can become a higher-risk code or policy change. The original classification does not survive material scope expansion.

Decision record

Preserve the task, dimensions, autonomy level, permissions, required evidence, reviewer, stop conditions, and final acceptance decision.

Sources

Follow the evidence.

  1. daltonanderson.net: how i built a go app in 4 hours with cursor aidaltonanderson.net
  2. owasp.org: www project top 10 for large language model applicationsowasp.org
  3. go.dev: getting startedgo.dev
  4. csrc.nist.gov: finalcsrc.nist.gov
  5. cursor.com: auto reviewcursor.com
  6. daltonanderson.ghost.io: how i built a go app in 4 hours with cursor aidaltonanderson.ghost.io
  7. cursor.com: 1 7cursor.com
  8. Spotify episodeopen.spotify.com
  9. cursor.com: teamscursor.com
  10. youtu.be: n4 J1tDwreMyoutu.be
  11. owasp.org: www project code review guideowasp.org
  12. cursor.com: privacycursor.com
  13. cursor.com: securitycursor.com
  14. slsa.dev: v1.2slsa.dev
  15. git-scm.com: v2git-scm.com
  16. cursor.com: agent best practicescursor.com
  17. cursor.com: data usecursor.com

From this episode

Two useful next steps.

Guide · 1 min

How to Write Project Rules for an AI Coding Agent

Write concise, scoped, versioned rules for AI coding agents, then move behavior, formatting, access, and release requirements into stronger enforceable controls.

Guide · 1 min

How to Review AI-Generated Code Before Accepting It

A rigorous review method for AI-generated code covering intent, scope, interfaces, tests, security, dependencies, operations, provenance, ownership, and rollback.

Return to the episode
AI Coding Task Autonomy Matrix