Closed feedback→behavior loop

Part of umbrella #10542 · Theme A (collaboration core) · highest-leverage latent asset

Problem

AutoBot captures everything needed to get measurably better at a given user/org over time — approvals, rejections, agent diary, user-scoped retrieval signals, replay logs — but the data is not yet wired into a loop that changes behavior. The agent does not learn "humans here keep rejecting my X, so shift routing/prompt/skill-selection." The hard part (capture) is done; the last wire (consume) is missing.

What exists today

• routers/feedback.py — feedback capture endpoint.
• rlm/evaluator.py + rlm/rag_refiner.py — within-response self-critique loop (Research: Recursive Language Model (RLM) patterns for AutoBot #1373).
• memory/agent_diary.py, memory/essential_story.py, trajectory store.
• knowledge/.../retrieval_learner* — user-scoped retrieval learning already exists.
• llc/models/replay_log.py + llc/api/replay.py — replay logs.
• Approval decisions flow through agent_loop/approval_workflow.py.

What to build

• A preference/feedback aggregator that turns captured approvals/rejections/edits/diary reflections into durable, queryable signals scoped per user + per org + per task-class.
• Consumption hooks: AgentRouter / StrategyPlanner / prompt_manager read those signals to bias routing, skill selection, and prompt assembly.
• Generalize the existing per-retrieval learner into a cross-cutting learning surface (don't duplicate — extend it).
• Guardrails: bounded influence, explainable ("adjusted because humans rejected Y 4×"), per-tenant isolation.

Acceptance criteria

• Repeated rejection of a behavior in a tenant demonstrably shifts the agent's next choice (test on a seeded feedback set).
• Adjustments are explainable and surfaced in the trajectory.
• Signals are tenant-isolated; no cross-org leakage.
• Extends existing retrieval_learner / diary rather than adding a parallel store.

Entry points

autobot-backend/routers/feedback.py · autobot-backend/memory/agent_diary.py · autobot-backend/knowledge/search_components/ (retrieval_learner) · autobot-backend/orchestration/agent_registry.py + enhanced_orchestration/workflow_planning.py (consumption) · autobot-backend/prompt_manager.py.