Agent Loop Skills 🧩 Skill Open source

What it does

Agent Loop Skills packages iterative improvement loops as open-standard Agent Skills. Each loop is generic: when you invoke it you bind it to your own task by naming the artifact to improve, the command that measures it, and a budget. The agent then proposes one change, runs it in your environment, scores the result against a real signal such as tests, latency or a metric, keeps the change only if it is better, logs it to an append-only ledger and repeats until a plateau, budget or threshold stops it. The author labels the project experimental and is candid that unsupervised loops can drift, which is why every loop is gated by an objective check.

What is inside

• Autoresearch: a minimal Karpathy-style loop plus analysis-first, exploratory, tournament, dueling and evolutionary variants for ML training code
• Code and optimization: optimize-loop for refactors or SQL speedups behind a correctness gate, prompt-optimize, and a plan-loop / swe-loop pair that turns a request into tasks and implements them with separate engineer and QA roles
• Research and writing: literature search over Semantic Scholar and arXiv, surveys, hypothesis generation, proposal grading, scientific writing and figures
• Data: hypothesis-driven analysis, anomaly investigation, claim verification and table cleanup
• Security: red, blue and purple team loops, explicitly for systems you own

Works with

The skills follow the Agent Skills format, so any compatible host can load them. Multi-role loops spawn real parallel subagents only on Claude Code; on Codex, Cursor and other hosts the same roles run inline, one after another.

How to install

npx skills add gaasher/agent-loop-skills

On Claude Code you can instead add the repo as a plugin marketplace and install agent-loops.

Maintenance and safety

The repo is MIT-licensed and was last updated in mid-2026. Loops shell out to commands you bind, so they can run code, train models and edit files for many iterations; set a budget and review what the ledger kept. Research loops call external paper APIs.

Who should use it

Developers and researchers who already have a measurable target, such as a test suite, a benchmark or a validation metric, and want an agent to iterate against it with a record of every attempt.