self-improving-agent-loop

self-improving-agent-loop is a skill for Claude Code, Codex from 45ck/skill-harness. It costs 39 tokens per session (338 once invoked), scanned A, original, MIT.

A guided cycle for improving an agent workflow over time. It collects evidence about failures, makes small reversible changes, checks them, and records supported lessons.

In plain words
What is it for?
Use it to improve prompts, skills, checkers, workflows, or operating rules. It helps collect issues and test results, choose one change, validate it, and track follow-up proposals.
Why use it?
It reduces guesswork when an automated workflow needs to improve. Human approval remains required for permissions, destructive actions, production data, publishing, merging, deployment, and policy changes.

Skill for Claude CodeCodex

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add skills/45ck/skill-harness/self-improving-agent-loop
Any agent
npx skills add 45ck/skill-harness --skill self-improving-agent-loop
Clone the repo
git clone --depth 1 https://github.com/45ck/skill-harness

Made for: Claude Code, Codex.

Per session 39 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 338 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00039 $0.00338
Opus 5 $0.00019 $0.00169
Sonnet 5 $0.00008 $0.00068
Haiku 4.5 $0.00004 $0.00034

Measured 3d ago against content hash c5c2dc57c46f, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

self-improving-agent-loop scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

packs/agent-operating-skills/skills/self-improving-agent-loop/SKILL.md · 37 lines

What it actually says

Self-Improving Agent Loop

Use this skill when a project wants agents to improve a workflow, skill, prompt, loadout, checker, or operating policy over time.

Loop

  1. Sense: collect issues, diffs, tests, traces, artifacts, user feedback, and handoff notes.
  2. Model: identify the task type, capability assumption, failure mode, and quality bar.
  3. Plan: choose one reversible improvement with explicit evidence.
  4. Act: change the smallest useful surface using existing project patterns.
  5. Gate: run tests, artifact checks, permission checks, and review gates.
  6. Learn: record a follow-up issue, durable memory, skill update, or checker proposal only when evidence supports it.

Governance

  • Treat frontier models as capable of planning, synthesis, review, and context work when the digital surfaces are available.
  • Keep deterministic scaffolding for repeatable validation, policy checks, manifests, and install wiring.
  • Treat generated traces and self-assessments as evidence candidates, not proof.
  • Require human approval for permission expansion, destructive actions, production data, publishing, merge, deployment, and policy changes.
  • Keep domain-specific loops out of the core pack unless the pattern generalizes across projects.

Output

Loop Goal

Sensors

Failure Model

Reversible Action

Gates

Human Approval Boundaries

Learning Output

Next Issue Or Memory

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 37 lines · 39 tokens per session scan A c5c2dc57c46f

Subscribe to this mod's changes

self-improving-agent-loop is a skill published in the GitHub repository 45ck/skill-harness (15 stars, last pushed 2mo ago), licensed MIT. It adds 39 tokens to every session and 338 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.

Related

Other skills, from other repositories

rubber-duck

Adversarial "rubber duck" review that turns explaining-out-loud into a hallucination check. The main session is the PRESENTER (it did the work — a design doc, investigation, or analysis — and holds the real reasoning) and reconstructs the topic to a LISTENER — a spawned subagent pinned to a DIFFERENT-vendor model that…

kirodotdev/KiroCrew · 161 tokens

llm-council

Convene a cross-vendor LLM council — the main session acts as Chairman and spawns several subagents, each pinned to a DIFFERENT model (Anthropic / OpenAI / DeepSeek / Zhipu / Qwen / etc. via kiro-cli). Three modes — synthesis (independent answers merged into one), vote (structured ballots + majority tally), and…

kirodotdev/KiroCrew · 209 tokens

context-retrieval-protocol

Context retrieval patterns for accessing session context via an MCP-compatible context server. Provides patterns for browse and retrieve, hybrid search, semantic search, and full-text search. Use when you need to retrieve previous context or search for relevant information.

alex-feel/mcp-context-server · 53 tokens

context-metadata-schema

Normative metadata schema v1 for all context-server records. Defines the kind registry (usermessage, report, plan, handoff, checkpoint, note, issue, comment), required and optional metadata fields per kind, status vocabularies, the typed links object with its design rules, and the filter recipes that make entries…

alex-feel/mcp-context-server · 134 tokens

context-preservation-protocol

Context preservation patterns for storing work results and session context via an MCP-compatible context server. Provides patterns for documenting work, storing reports, and ensuring continuity between sessions. Use when you need to preserve work results or session context.

alex-feel/mcp-context-server · 50 tokens

issue-tracking-protocol

Cross-project task tracker hosted on the context server. Use whenever you need to file an issue, bug report, feature request, improvement, or follow-up task for ANY project (your own or another); triage, prioritize, assign, or transition your own project's issue through its lifecycle; pick up an issue in your own…

alex-feel/mcp-context-server · 179 tokens