Adaptive Auto-Harness

Code for the paper "Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams."

📦 This is a mirror. The canonical, actively-maintained code lives in the release/adaptive-auto-harness branch of the A-Evolve project — the universal infrastructure for self-improving agents. Star and follow development there:

• 🌱 Release branch (this paper): a-evolve/tree/release/adaptive-auto-harness
• 🧬 A-Evolve main (the framework): a-evolve/tree/main

Project Structure

.
├── agent_evolve/                  # Core library
│   ├── algorithms/                # Evolution engine + routing + adaptation
│   ├── agents/{polybench,ctf_dojo,futurex}/   # Task-solving agents
│   ├── benchmarks/{polybench,ctf_dojo,futurex}/  # Task loading + scoring
│   ├── engine/                    # Evolution loop, versioning, observer
│   ├── protocol/adaptation/       # Pluggable solve-time adaptation operators
│   ├── contract/, llm/, tools/, utils/
├── experiments/                   # Configs + evolver prompts + seed harness
├── seed_workspaces/               # Initial harnesses per benchmark
├── scripts/                       # poly_/ctf_dojo_/futurex_hypothesis.sh launchers
├── evaluations/analysis_poly/     # Scripts that regenerate the README figures
├── assets/                        # README figures
├── data/                          # Dataset fetch helper + layout notes
└── solve_all_with_evolution.py    # Main entry point (all benchmarks)

Installation

Requires Python 3.11+ and (for PolyBench) AWS Bedrock access.

git clone -b release/adaptive-auto-harness https://github.com/A-EVO-Lab/a-evolve.git
cd a-evolve
conda create -n aevolve python=3.11 -y && conda activate aevolve
pip install -e ".[all]"

Running PolyBench

PolyBench is pure reasoning (no Docker). From the repo root:

# 1. Configure credentials + models (export into the shell; scripts read env vars)
cp .env.template .env        # then edit: SOLVER_MODEL, EVOLVER_MODEL, AWS_* 
set -a; source .env; set +a

# 2. Get the dataset (SQLite snapshot of Polymarket markets)
python data/download_data.py --benchmark polybench   # -> data/polymarket_analysis.db

# 3. Smoke test: no-evolution baseline on 5 markets
bash scripts/poly_hypothesis.sh --limit 5 H0

Then run the paper's hypothesis cells (omit the target to run all):

bash scripts/poly_hypothesis.sh H0            # baseline: no evolution
bash scripts/poly_hypothesis.sh H1            # full evolution
bash scripts/poly_hypothesis.sh H4_multi      # multi-agent structured evolution
bash scripts/poly_hypothesis.sh H4_multi_nav  # + tree routing

Add --adaptation <name> to select the solve-time operator (default: tree_routing when routing is enabled, else whole_store). Results land in results/polybench_<cell>/, logs in logs/.

CTF-Dojo and FutureX run the same way via scripts/ctf_dojo_hypothesis.sh and scripts/futurex_hypothesis.sh. See INSTALL.md for the full provider matrix and dataset notes.

Pluggable Adaptation

Solve-time adaptation is a pluggable operator chosen with --adaptation <name>; operators live in agent_evolve/protocol/adaptation/.

--adaptation	Granularity
whole_store	full harness
tree_routing	whole branch (agentic router, paper default)
retrieval	per-task top-k
agentic_filter	per-task LLM-selected subset

More operators are on the TODO list (e.g. graph-structured store, lazy loading, dependency-aware retrieval) — contributions welcome: add a class in operators.py and one line in registry.py.

Citation

If you find this work helpful, please consider to cite our paper:

@article{liu2026adaptive,
  title={Adaptive Auto-Harness: Sustained Self-Improvement for Agentic System Deployment on Open-Ended Task Streams},
  author={Liu, Zewen and Shi, Zhan and Sang, Yisi and He, Bing and Lin, Minhua and Wei, Tianxin and Wang, Dakuo and Dumoulin, Benoit and Jin, Wei and Lu, Hanqing},
  journal={arXiv preprint arXiv:2606.01770},
  year={2026}
}

License

MIT.