Skip to content

INSTALL.md

Latest commit

 

History

History
89 lines (65 loc) · 3.18 KB

INSTALL.md

File metadata and controls

89 lines (65 loc) · 3.18 KB
Raw

Installation

Requirements

  • Python 3.11+
  • Docker (for sandboxed evolver and CTF-Dojo task containers)
  • 50+ GB free disk for benchmark data and trajectories
  • One supported LLM provider (see below)

Install

Unpack the supplement archive (e.g., tar -xzf adaptive-auto-harness.tar.gz), enter the resulting directory, and install the package in editable mode:

cd adaptive-auto-harness
pip install -e ".[all]"

The [all] extra pulls every supported LLM provider and benchmark backend. To install a minimal subset, replace [all] with one or more of: [bedrock], [openai], [litellm], [swe], [mcp], [skillbench], [gepa].

Environment variables

Copy .env.template to .env and fill in the values that apply to your setup. The framework reads .env automatically.

Required

Variable Purpose
SOLVER_MODEL Model ID used by the task-solving agent
EVOLVER_MODEL Model ID used by the evolver agents

The placeholder strings <solver-model-id> and <evolver-model-id> appear throughout the codebase. They are resolved at runtime from the above env vars; replace them before running, or set the env vars and the loader will substitute.

Provider credentials (one of)

Provider Required env vars
AWS Bedrock AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, AWS_REGION
OpenAI OPENAI_API_KEY
LiteLLM (any backend) provider-specific (see https://docs.litellm.ai/)

The default LLM client is in agent_evolve/llm/. To swap providers, edit the BedrockClient or add a new client class implementing the same interface; configs reference clients by name.

Optional retrieval API keys (FutureX live mode)

Variable Purpose
TAVILY_API_KEY Web search (Tavily)
FRED_API_KEY Economic indicators (FRED)
SERPER_API_KEY Web search (Serper)
JINA_API_KEY Web search (Jina)
EXA_API_KEY Web search (Exa)

Without these, FutureX runs in strict mode using only Wikipedia revisions plus DuckDuckGo with date filtering.

Data

Datasets are not shipped with this repository. From the project root:

python data/download_data.py --benchmark polybench    # ~40 MB
python data/download_data.py --benchmark ctf_dojo     # ~2 GB
python data/download_data.py --benchmark futurex      # ~10 MB

See data/README.md for the source of each benchmark and any manual prerequisites (e.g., CTF-Dojo Docker images).

Verifying installation

python -c "import agent_evolve; print(agent_evolve.__file__)"
pytest tests/ -q

A successful test run validates the framework imports, contract helpers, and a small evolution dry-run.