Automatic
- Research diffs and freshness checks
- Drafts and test execution
- Evidence collection and internal reports
Operate · supervised autonomy
Let agents run the routine. Keep authority explicit.
One recurring loop observes the system, prepares changes, verifies evidence, and escalates only the decisions that require a responsible person.
Weekly system loop
Observe → decide → verify → learn.
The output is a decision record and one bounded experiment, not another stream of chat.
- 01
Observe
Collect customer, delivery, reliability, cost, and market signals.
- 02
Score
Rank exceptions and opportunities against explicit criteria.
- 03
Propose
Write the next action, expected evidence, and stop condition.
- 04
Validate
Run policy, source, test, and authority checks before action.
- 05
Execute
Perform only the actions permitted for the current lane.
- 06
Verify
Record the output, acceptance evidence, cost, and residual risk.
- 07
Learn
Update the system contract, knowledge, or experiment decision.
Authority lanes
Autonomy increases only with evidence.
Every capability belongs to a visible lane. Higher financial, legal, privacy, or reputation impact means stronger approval.
Review before action
- Customer-facing deliverables
- Partner destinations and disclosures
- Material scope or workflow changes
Principal approval
- Contracts, pricing, and public income claims
- Production promotion and credential grants
- Payments, transfers, wallets, and bank actions
Run artifact
income/weekly/YYYY-WW.md
Decision budget
One accepted experiment per loop
Escalation rule
Only authority, exception, or material-risk decisions