BioAgent Gym is a repository of independently installable biomedical agents and native Harbor leaderboards. Harbor 0.23.0 owns Docker environments, trials, timeouts, verification, and result storage. This repository supplies agents, dataset tasks, scoring code, and native job YAML.

The supported catalog is agents/README.md. Install only the shared item runner and the agent you intend to use:

python3.12 -m venv .venv-harbor
.venv-harbor/bin/pip install -e . -e runner -e 'agents/dswizard[harbor]'
.venv-harbor/bin/pip install -e 'agents/deepevidence[harbor]'

Coder, DSWizard, and DeepEvidence are separate distributions. Their wheels own only their agent namespace; bioagent-harbor-runtime is a separate lightweight distribution.

The current leaderboard catalog is benchmarks/README.md:

• BioDSBench
• Biomedical Deep Research

Each benchmark directory contains its tasks, jobs, scoring source, preparation script, and tests. Prepare pinned data once:

.venv-harbor/bin/python benchmarks/biodsbench/prepare.py
.venv-harbor/bin/python benchmarks/biomedicine-deep-research/prepare.py

Preparation writes public items, private verifier references, large external inputs, and grader deployment copies. It never rewrites committed task.toml, Dockerfiles, instructions, or job YAML.

Run deterministic smoke jobs:

ROOT="$(pwd)"
.venv-harbor/bin/harbor run \
  -c benchmarks/biodsbench/tests/jobs/dswizard-mock.yaml \
  --jobs-dir "$ROOT/.harbor/jobs" -y

.venv-harbor/bin/harbor run \
  -c benchmarks/biomedicine-deep-research/tests/jobs/deepevidence-mock.yaml \
  --jobs-dir "$ROOT/.harbor/jobs" -y

Run the formal leaderboard configurations with the required provider secret:

.venv-harbor/bin/harbor run \
  -c benchmarks/biodsbench/jobs/dswizard.yaml --env-file .env \
  --jobs-dir "$ROOT/.harbor/jobs" -y

.venv-harbor/bin/harbor run \
  -c benchmarks/biomedicine-deep-research/jobs/deepevidence.yaml --env-file .env \
  --jobs-dir "$ROOT/.harbor/jobs" -y

.venv-harbor/bin/python benchmarks/biomedicine-deep-research/summarize.py \
  "$ROOT/.harbor/jobs/<deepevidence-job-directory>"

BioDSBench ranks the 118-item Python task; R remains inventoried but unsupported. Biomedical Deep Research ranks verifier-split choice items by exact match and evidence-gap retrieval by recall@30. To experiment on fit or tune, copy the formal YAML into the gitignored local-jobs/ directory and change both the agent kwargs.split and verifier BIOAGENT_SPLIT; these values must agree.

Use harbor view .harbor/jobs to inspect submissions, item artifacts, verifier logs, and rewards. See architecture.md, adding-agents.md, and adding-benchmarks.md. Historical code is isolated under legacy/.