Triple Loop Learning

---
name: triple-loop-learning
plugin: agent-orchestration
description: "(Industry standard: Meta-Learning System / Automated Autoresearch) Autonomous improvement loop evaluating friction and validating mutations against headless benchmarks."
allowed-tools: Bash, Read, Write
---

## Dependencies

This skill requires **Python 3.8+** and standard library only.

**Evaluation gate**: NOT included in this primitive. The calling system (e.g., agent-agentic-os os-improvement-loop) is responsible for wrapping this skill with an eval gate and experiment log.

---

# Triple-Loop Learning (`triple-loop-learning`)

Autonomous multi-session improvement architecture that identifies friction, forms hypotheses, and tests mutations against headless benchmarks.

## Contents

- [Dependencies](#dependencies)
- [Constraints](#constraints)
- [Quick start](#quick-start)
- [Architecture](#architecture)
- [Workflow](#workflow)
- [Verification](#verification)
- [References](#references)

## Constraints

- **Objective scoring**: Subjective self-evaluation is prohibited; acceptance requires deterministic tests and score differentials.
- **Process suspension**: Always append `< /dev/null` to background sub-agent execution.
- **Promotion gate**: Only promote mutations where regression suites pass and scores exceed the established baseline.

## Quick start

```bash
pytest plugins/agent-orchestration/tests/test_loop_strategies.py
```

## Architecture

```mermaid
flowchart TD
    subgraph Outer["Meta-Learning"]
        Hypothesize --> StrategyBridge[Strategy Packet]
        Report --> Conclude[Accept / Reject]
    end
    subgraph Mid["Planner"]
        Plan[Define Sub-tasks] --> TacticalBridge[Handoff]
        Result --> Report[Score Analysis]
    end
    subgraph Inner["Executor"]
        Execute[Mutation] --> Test[Headless Eval]
        Test --> Result
    end
    StrategyBridge --> Plan
    TacticalBridge --> Execute
```

## Workflow

1. **Friction Ingestion**: Ingest logs and cluster repeated friction events.
2. **Hypothesize**: Formulate testable hypothesis ("Modifying X improves metric Y").
3. **Dispatch**: Select backend, author strategy packet, and assign tasks.
4. **Mutate & Score**: Tactical executor mutates code and runs headless tests.

## Verification

```bash
pytest plugins/agent-orchestration/tests/test_loop_strategies.py
git diff --check
```

## References

- [acceptance-criteria.md](references/acceptance-criteria.md) — Acceptance criteria for autonomous improvement cycles.
- [fallback-tree.md](references/fallback-tree.md) — Fallback escalation protocol for loop stagnation.