Adaptive behavior v1: learn from feedback and prove one useful improvement

[JRichlen]

Build an agent that improves how it helps a person over time. Start with review: adapt depth, domain lenses and output to the user's goal. Extend to learning and coding after demonstrating benefit.

Lifecycle: understand → select behaviors → act → collect feedback → compare with baseline → approve or reject → monitor and roll back. Preferences remain scoped, inspectable, correctable and deletable.

Now

Control mechanisms have been exercised; user benefit and judge calibration remain unproven. Runtime expansion is paused. Detailed lab evidence stays private.

Next: one useful experiment

• Freeze a small corpus around “issues contain too much context.” Set thresholds for less editing and easier comprehension without losing required facts or actions.
• Compare baseline and one concise-review candidate on held-out tasks. Retain outputs, deterministic checks, blinded human labels and total usage. Calibrate an independent frontier judge against human labels; external grading requires separately authorized inputs and spending limits.
• Report improvement, regression or inconclusive evidence. Only a supported result advances to an approved, scoped install with demonstrated rollback.

Fit and boundaries

Reuse agent-compiler for composition, recurrence-detector for candidate signals, and existing verification, packaging and curation patterns. Evaluate the framework in #115 and calibration fix in #114 before adding another harness. Both remain unqualified for behavioral benefit.

Actor qualification follows the separate upstream plan: pinned local Qwen 27B, direct 32K/one-sequence baseline, then measured context lanes; dispatcher continuity and Aperture follow their own acceptance gates. Capacity is unqualified, AgentWorld is parked, and the actor route has no external fallback. The request contract stays in #101/#107; this plugin consumes that capability.

Later

Learning/coding transfer and opt-in curation automation. First run should preview a natural-language schedule, scope and limits, then deploy only after approval; repeated setup must not duplicate it, and pause/removal must work. Trigger semantics (#84), broader testing architecture (#89), and the optional cost benchmark (#116) are deferred references, not additional active workstreams.

Tracking: update this checklist at meaningful checkpoints. Preserve historical discussions (#85, #102); keep protocols in artifacts and avoid child-ticket sprawl.