Preferences update
ALLOW. The intended account preferences patch executed. No human intervention.
Illustrative evidence / sample data
An Argos
evidence record.
This static engineering example is not a database-backed Task. It illustrates how Argos links an objective, controlled operations, human decisions, execution results, and independent verification.
Task / account-settings
Add account notification preferences.
- Declared scope
- Expected: src/accounts/preferences.ts
Review: src/shared/settings.ts
Protected: src/auth/session.ts - Agent
- Codex 0.143.0 / supported apply_patch updates
- Execution
- Two controlled updates recorded. One protected proposal blocked before mutation.
- Verification
- VERIFIED
- Evidence source
- Independent Git/filesystem readback and the frozen deterministic checks.
Recorded operations
What was allowed. What actually happened.
Shared settings update
Held before mutation. An authenticated human approved the exact pending operation, which then executed.
Session update
BLOCK. The proposed authentication change was not executed. No repository mutation from this operation.
Independent verification
The conclusion has a declared basis.
- Required expected change observed and consistent with controlled evidence
- Protected paths unchanged
- Dependency integrity unchanged
- Typecheck passed
- Focused tests passed
- Build passed
VERIFIED
Every required condition in this illustrative frozen plan passed. A change-required Task with no corresponding repository outcome would be FAILED, even if its commands passed.
If repository truth or a required executable were unavailable, verification would be UNCERTAIN. Execution history would remain separate.
Real Task and Evidence history
The URL is a destination, not a credential.
Real Task and Evidence links are authentication-protected. Login returns an authorized user to the requested record. Possession of a URL does not grant access.
Your next engineering Task
AI is getting more capable.
Your supervision model should too.
Start with Codex, one repository, and a boundary you can inspect.