The idea
Most agents have a fixed toolbox. When a request falls outside it, they either fail or improvise a one-off answer that is forgotten afterwards.
Gara answers a request by finding a skill for it, or building one. The next time a similar request arrives, retrieval hits and none of the building has to happen again.
How it works
1. RETRIEVE hit / adapt / miss
2. ADAPT a skill is close: upgrade or fork it
3. TRIAGE does this need a skill? split it into steps
4. CLARIFY is the requirement clear? if not, ask
5. PLAN run independent steps in parallel
6. FEDERATE search the marketplace and install
7. RESEARCH web search + fetch
8. SPECIFY draft the spec and eval cases, audit them
9. AUTHOR write the code
10. EVALUATE run the evals in a sandbox
11. PROBE attack it with input it never saw
12. CRITIQUE review for over-reach and correctness
13. PROMOTE install, embed, index
14. EXECUTE → VERIFY → SHARE (human-approved)What’s interesting
- It improves what it has instead of starting over: retrieval can hit, adapt an existing skill, or miss.
- Every skill ships with its own eval suite, so “it works” is checked, not assumed.
- A critique step reviews each new skill for over-reach before it is allowed to install.