Inform
Research and explain. Nothing outside the conversation changes.
The working loop
Six steps, with evidence at the end.
A request is not complete when a model has produced an answer. It is complete when the requested result exists and has been checked against the real system.
- 01Request
Start with the outcome the human actually wants, not the tool that first comes to mind.
- 02Context
Load only the relevant conversation, durable preferences, reusable procedures and live source data.
- 03Decision
Choose between direct reasoning, a deterministic script, a tool call or a focused subagent.
- 04Action
Read before writing, keep the scope narrow and preserve a way back when the change is risky.
- 05Verification
Test the artifact, inspect the external result and treat success messages as claims rather than proof.
- 06Retention
Save stable preferences as memory and repeatable methods as skills. Leave temporary task state behind.
Autonomy ladder
Freedom is earned one reversible step at a time.
The same agent can operate at different levels. The right level depends on impact, reversibility and how clearly success can be measured.
Lowest impact
Propose
Prepare a concrete change, but let the human execute or approve it.
Act with approval
Do the reversible preparation, then stop before the consequential step.
Act and verify
Complete a bounded task, run the checks and report the evidence.
Run on schedule
Operate unattended only behind deterministic gates, rollback and clear failure reporting.
Three kinds of context
Remember less, but remember the right things.
Session
The active job, its temporary state and the evidence gathered along the way.
Expires with the workMemory
Stable preferences and facts that should prevent the human from repeating themselves.
Compact and durableSkill
A tested procedure with exact steps, pitfalls and verification that can be reused later.
Improves with practiceHuman checkpoints
Some decisions should stay human.
- Irreversible changesDeletion, publication or changes that cannot be rolled back safely.
- Unclear intentWhen two reasonable interpretations lead to meaningfully different outcomes.
- Identity and judgementMessages, commitments or decisions where the human should remain the author.
- New trust boundariesCredentials, broader permissions or a new system receiving private data.