How agents adapt and learn to act.
Self Research studies how agents adapt to people and businesses, learn when to act proactively and earn more autonomy.
What we’re working on.
Learning from corrections: What should one correction teach an agent?
Each correction becomes a scoped rule, tested on past work and weighed by the time it saves.
Support agents that need each correction once.
Memory you can trace: Can an agent say where it learned each fact?
Every memory keeps its source; unsourced facts never act.
Any decision an auditor asks about, traced to who taught it.
Skills on smaller models: Can a small model do what a large one learned?
We distill repeated work into skills that small models pass.
IT tickets closed at large-model quality, at small-model cost.
Knowing when to ask: When should an agent stop and ask?
We check confidence against outcomes and set the line by cost.
A month-end close that asks only when being wrong costs more.
Autonomy that is earned: How does an agent earn the right to act alone?
Trust grows with verified work and resets when the model changes.
Revenue agents trusted to sign renewals, on their own record.
Results checked in the record: Did the work actually happen?
Outcomes are confirmed where they live: the ledger, the ticket, the repo.
Incidents closed only once the dashboards recover.