Closed learning loop
Okami gets better with use. A closed learning loop watches every run — steps, tools, errors, gates and your feedback — and writes back into the agent: skills, memor…
The loop
runs ─► logs trajectory (steps, tools, errors, gates, feedback)
▲ │ post-task reflection
│ creates/updates a SKILL (via skill.sh, scanned)
│ writes a FACT / ANTI-PATTERN to memory
│ tunes the model's capability profile
│ proposes PERSONA / VOICE evolution
applies updates the design TASTE modelAuto-skill
With learning.auto_skill: true, the agent distills a skill from every non-trivial task and installs it through the same path as okami learn — passing the security scan before going live (agent-created skills are no exception). What it learns to do well becomes reusable.
Anti-patterns
A recurring failure isn't forgotten: it becomes a negative rule (guard-rail) in memory. Next time, the harness injects the anti-pattern into context — “I already tried X and it failed because Y; don't repeat”.
Capability-profile auto-tune
The loop measures per-model failure modes (gate rejections, stalls, malformed tool-calls) and adjusts the profile: more decomposition, stricter grammar, more retries, smaller subsets. Each model gets a scaffold tuned ever better to it. See okami tune.
Curation and governance
Periodic curation (reflect) consolidates and prunes rotten memory, avoiding context-rot. Sensitive changes (a permanent persona, a new skill) can require human approval — everything is versioned and reversible (/undo, append-only changelog).