Overview
AutoHarness examines how Claude Code can turn corrections from real work into bounded, inspectable procedural memory. The series follows the path from tool activity and redacted episode capture through model proposals, queued deterministic promotion, scoped recall, use/view/patch signals, lifecycle pruning, and whole-library curation. Its central thesis is that memory is useful only when learning, persistence, discoverability, ownership, and retirement are governed together. The articles support a coherent maintenance design with provenance, reversible archives, and representative defenses, but not a claim that accepted rules are correct, improve task outcomes, or work reliably across every live host environment.
Learning outcomes
- AutoHarness treats agent memory as a maintenance loop, not a larger store: lessons must be captured, admitted, surfaced, measured, and retired.
- A model may propose a lesson, but the queued intent and deterministic promoter—not the reflector—control changes to live skills.
- Use, view, and patch signals describe adherence, recall value, and maintenance separately; none proves task correctness.
- The design supports bounded, inspectable procedural continuity, but its evidence does not establish better outcomes or universal host reliability.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What problem does AutoHarness solve?
AutoHarness addresses two failures of agent memory: lessons disappear with the session, or persistent rules accumulate until they are hard to find and trust. It aims for a smaller, inspectable library rather than memory growth alone.
How does AutoHarness turn a correction into a reusable skill?
AutoHarness captures a bounded, redacted work episode, has a reflector propose a lesson, queues the intent, and lets a deterministic promoter decide whether it can land. Later recall and maintenance provide feedback.
How does AutoHarness keep model-generated lessons from changing files directly?
The reflector cannot write the live skill tree: `stage_skill` records a structured intent, and the promoter rechecks scope, structure, safety, ownership, and paths before writing. A rejected proposal should leave no live change.
Can AutoHarness prove that a learned skill improves task outcomes?
No. Its tests and demonstration establish mechanisms such as queue behavior, recall, lifecycle, and representative defenses, but maintenance signals do not show that a rule improved the resulting task. Live end-to-end behavior also depends on the Claude Code host.