self-improvement-loops

What it does

The skill enables systems where the harness or optimizer itself is the artifact to be improved. It covers recursive self-improvement loops, meta-harnesses, self-improving harnesses that mine failures and propose bounded edits, evolutionary or population-based search over agent scaffolds, and acceptance gates for self-modifying systems. It directs governance of a single autonomous loop with bounded surfaces, durable logs, rollback, novelty gates, and approval boundaries, aimed at harness-engineering, evaluation, and hosted infrastructure management.

How it works

The skill instructs designing and operating self-modifying cycles that mine failures, generate bounded edits to surfaces they can modify (prompts, context playbooks, workflow graphs, harness code, and potentially the optimizer), and then evaluate those edits using a two-split acceptance gate (held-in and held-out splits) before merging or rejecting changes. It prescribes maintaining an external, auditable archive of candidates and traces, avoiding summarization within the prompt, and ensuring that escalation, rollback, and governance remain outside the mutable surfaces. It advocates meta-level search over harness code, context evolution with versioned mechanisms, and the separation of evaluator details from the proposal signal. It also outlines practical readiness checks, escalation criteria, and guidance for choosing the appropriate rung of the optimization ladder when addressing failures.

When to use it

Activate this skill when: you are building a loop where an agent proposes edits to its own harness or workflow based on mined failure patterns; you are evaluating meta-agent search over harness code; you need to set acceptance criteria for self-modifying systems; you are evolving the context mechanism itself; you are diagnosing degenerating self-improvement loops; or you need to decide which level (prompt, context, workflow, harness, or optimizer) to fix a recurring failure.

What it can touch

The skill explicitly references editable surfaces such as prompts, context playbooks, workflow graphs, harness code, and the optimizer code. It prescribes an archive layout containing source, scores, traces, and lineage, and requires a two-split acceptance gate for edits. It emphasizes keeping evaluation logic and permission checks external to the surfaces the loop can modify, and storing raw traces for auditing rather than summarizing them in prompts.

Caveats

It specifies that the evaluator, permission controls, and budgets live outside the mutable loop, and that any exploit or regression should be logged with evidence. It requires a capable, held-out split and a deterministic, automatable evaluator, plus an auditable history for diffs and traces. It warns that the loop may not be suitable for domains with slow or ambiguous evaluation, and that human decisions remain at points for evaluator changes or surface expansions. The license is MIT.