SkillOptLite and HarnessOpt: Better and Faster Agent Self-evolution with One Line of Vibe

Train your agent's skills and harness with a single line of vibe — automatic, looped improvements.

👉 Project Page 👈

👉 WeChat Group 👈

---

TL;DR

Open this repo folder inside your coding agent (VS Code Copilot Chat, Codex CLI, Claude Code, kimi-code, glm-code, deepseek-tui — anything that reads .github/prompts/*.prompt.md), then type the commands below into that same coding-agent chat — not a plain shell. The coding agent handles environment setup for you (pip install, auth, data download); you never leave the chat.

1. Optimize a skill on a shipped benchmark (edit skill.md only). Open copilot_example/livemath (or spreadsheetbench / alfworld / docvqa / officeqa / searchqa) with your coding agent, then type at the chat prompt:

/skillopt-loop rounds=10 batch=40 target=gpt-5.4-nano

2. Co-optimize skill + agent harness (also edit the Python code). Open harness_example/spreadsheetbench with your coding agent, then type at the chat prompt:

/harnessopt-loop rounds=2 batch=40 target=gpt-5.4-nano skill=skill_best_nano.md

3. Bring your own repo + data — no shim to write; point the coding agent at a reference env and let it scaffold. Sample vibe:

Read copilot_example/livemath/ as a reference (env code + .github/prompts/skillopt-loop.prompt.md). Build a new env at copilot_example/myrepo/ that runs my eval: python -m myrepo.eval --items X --skill Y --out results.jsonl. My raw data is a single data/all.jsonl — split it 2:2:6 (train / val / test) with a fixed seed. Each result row has {id, input, expected, predicted, success}. Also copy and adapt the slash-command prompt files to my layout — skillopt-loop.prompt.md for skill-only, and harnessopt-loop.prompt.md (see harness_example/spreadsheetbench/.github/prompts/) if I want harness co-optimization; rewrite their split sizes, allow-listed editable files, and rollback tag prefix to match my directory. Smoke-test with --eval_limit 5 and confirm samples/failed/*.md is non-empty.

Once it reports done, run flow #1 or #2 against the new folder.

In all three cases the coding agent drives the loop — rollouts, sample inspection, skill.md patches, val-gated keep-or-revert, archive to workspace/.skillopt/history/. You don't touch anything else. When it stops improving, workspace/skill.md is the artifact you ship.

Two slash commands, one loop:

Command	Cwd it expects	What the coding agent is allowed to edit
/skillopt-loop	inside copilot_example/<env>/	just skill.md
/harnessopt-loop	inside harness_example/<env>/	skill.md and the agent code (rollout.py, react_agent.py, codegen_agent.py, executor.py, adapter.py)

Both are just prompt files at <env>/.github/prompts/*.prompt.md — read them to tune the loop policy (round count, gate discipline, dead-band, roll-back tag names).

Pipeline

Performance

A few concrete headlines from the release ckpts:

Setup	Score
SpreadsheetBench, GPT-5.4-nano + HarnessOpt (skill_best_nano.md)	0.7758
SpreadsheetBench, GPT-5.5 baseline harness + full SkillOpt	0.7620
LiveMath, GPT-5.5 + SkillOptLite	+8.8 pts over SkillOpt
LiveMath, GPT-5.4-nano + SkillOptLite	+25.4 pts over SkillOpt
ALFWorld, GPT-5.4-nano + SkillOptLite	81.3 (+9.5 over SkillOpt)
Spreadsheet, avg + SkillOptLite	+12.6 pts over SkillOpt

All numbers come from the ckpts in this repo — reproduce them with scripts/eval_only.py (see Direct eval below).

---

What's in this repo

copilot_example/                  ← 6 benchmarks, run.sh + /skillopt-loop each
├── searchqa/    livemath/    alfworld/
├── docvqa/      officeqa/    spreadsheetbench/
├── env.sh                        thin auth delegator
└── make_samples.py               results.jsonl → per-item .md for the loop

harness_example/                  ← full harness training loop
└── spreadsheetbench/             baseline code + starting skills + /harnessopt-loop

skillopt_lite_ckpt/               ← baked skills produced by /skillopt-loop
└── <env>/<model>/skill.md        (per env × per target model)

harnessopt_ckpt/                  ← end-state of /harnessopt-loop
└── spreadsheetbench/             optimized code + best_skill_*.md snapshot

skillopt_ckpt/                    ← paper GPT-5.5 reference skills
└── <env>/gpt5.5_skill.md

skillopt/  scripts/  configs/     ← runtime the run.sh scripts call into
data/                             ← download.py + local split cache
docs/                             ← quickstart, auth, data
env.sh  .env.example              ← 3-mode auth bootstrap

Datasets are hosted on 🤗 yshenaw/SkillOpt_Lite_Benchmarks.

---

Setup (fallback — your coding agent normally handles this)

Skip this section unless you're running headless or want to prep the environment before opening the coding agent. In the normal flow, the first time you issue any TL;DR command your coding agent will pip install -r requirements.txt, prompt you to fill in .env for whichever auth mode fits (az login / OpenAI key / OpenAI-compatible endpoint), source ./env.sh, and python data/download.py before starting the loop.

Manual bootstrap:

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env && $EDITOR .env       # AZURE_OPENAI_ENDPOINT / OPENAI_API_KEY / ...
source ./env.sh
python data/download.py                     # ~a few hundred MB

Full auth walkthrough (Azure CLI / Azure key / OpenAI / OpenAI-compatible) in docs/authentication.md.

---

Roadmap / TODO

• Agent-agnostic loop runner. A drop-in script that wraps /skillopt-loop around any existing agent codebase (not just the copilot_example/ and harness_example/ layouts). Bring your own rollout(item) → trace, score and a skill.md entry point, get the same round-based train → patch → val-gate → rollback loop. (Coming soon.)
• Codex CLI plugin. Port the /skillopt-loop and /harnessopt-loop slash-commands to OpenAI Codex CLI so the same loop works outside VS Code.
• Claude Code plugin. Same, for Anthropic Claude Code — package the prompts as a slash-command extension.

Contributions welcome — open an issue at EvolvingLMMs-Lab/SkillOpt-Lite if you want to help land any of the above.

---

Cite

@article{shen2026skilloptlite,
  title  = {SkillOpt-Lite: Better and Faster Agent Self-evolution with One Line of Vibe},
  author = {Shen, Yifei and Li, Bo and Zhang, Xinjie},
  year   = {2026},
  note   = {arXiv link coming soon}
}

---

Contact

• WeChat community: scan the QR above
• Correspondence: yshenaw@connect.ust.hk
• Issues / PRs: welcome at EvolvingLMMs-Lab/SkillOpt-Lite

Acknowledgements

Built on top of SkillOpt — the text-space optimizer that trains reusable natural-language skills for frozen LLM agents through trajectory-driven edits, validation-gated updates, and deployable best_skill.md artifacts. SkillOpt-Lite is the minimal coding-agent-driven variant of that loop; HarnessOpt extends the same loop to also edit the agent code. Huge thanks to the SkillOpt authors and contributors for the original design and open-source release.

Released under the MIT License.