Two ways to improve an AI operator
How expert knowledge and self-improvement can help build autonomous operators, and why ongoing responsibility for business outcomes needs clear standards and review.
- Harness²
- Source of truth
- Cost per item
Methodic Labs / Our Research
Notes from building
operators that do real work
Technical posts on how we build, evaluate and deploy operators, including what did not work.
2 postsLatest post October 15, 2026
Posts
Newest firstComparison
System design
Harnessing Autonomous Intelligence
How Refinery turns institutional knowledge into autonomous operators that own business outcomes, with expert rules, evaluations and autonomy earned through review.
- Data ingestion
- Rule mining
- Evaluation
The questions behind the posts
Company-specific intelligence needs company-specific tests
Explore Refinery01 / Context
What was knowable at the time?
Represent the records, policy versions, and unresolved facts available when a decision was made.
02 / Control
What is the system allowed to change?
Make action permissions explicit, enforce them outside the model, and test when the agent should defer.
03 / Evidence
Does it beat a strong baseline?
Compare against the same model with the client’s own material, on held-out cases.