AgentBus

AgentBus is a lightweight experimental runtime for low-overhead multi-agent communication, structured state transfer, and shared memory reuse.

The public repository is intentionally code-first. Large model files, benchmark datasets/results, third-party source trees, paper drafts, and submission templates are kept local and ignored by Git.

What Is Included

• agentbus/: core protocol, memory, mock runtime, and shared script helpers.
• scripts/: reproducible experiment and scoring entry points.
• configs/: lightweight schemas and runnable configuration stubs.
• tests/: standard-library unit tests for the core runtime and scripts.

What Is Not Vendored

• models/
• experiments/results/
• third_party/
• docs/
• latex/
• local benchmark datasets and generated artifacts

Some optional ALFWorld/Interlat adapter scripts still exist as code, but the upstream Interlat repository and model weights are not bundled. Core tests and the ScienceWorld scoring entry point run without vendored third-party source.

Quick Test

python -m unittest discover -s tests -v

Mock Runtime

python scripts\run_mock_experiment.py --rounds 10

Outputs are written under experiments/results/, which is ignored except for a directory placeholder.

ScienceWorld Real Simulator Scoring

Create or activate an environment with Python 3.8+, Java/OpenJDK, and scienceworld installed. The project-local environment used during validation was:

conda create -p .\.envs\scienceworld python=3.8 openjdk=11 pip -c conda-forge -y
conda run -p .\.envs\scienceworld python -m pip install scienceworld==1.2.3

Run a small real simulator score:

$env:PATH = "$PWD\.envs\scienceworld\Library\bin;$PWD\.envs\scienceworld;$PWD\.envs\scienceworld\Scripts;$env:PATH"

.\.envs\scienceworld\python.exe -u scripts\run_scienceworld_real_score.py `
  --task-nums 0,3,13 `
  --variations 0 `
  --policies gold_path,random_valid `
  --simplification easy `
  --max-steps 100

Run the 30-task variation-0 slice:

.\.envs\scienceworld\python.exe -u scripts\run_scienceworld_real_score.py `
  --task-nums all `
  --variations 0 `
  --policies gold_path,random_valid `
  --simplification easy `
  --max-steps 100

The scorer writes scores.csv, aggregate.csv, metadata.json, and summary.md under experiments/results/scienceworld_real_score/<run_id>/.

Automatic Version Push

This workspace can use the tracked Git hooks in .githooks/:

git config core.hooksPath .githooks

After each commit, the post-commit hook creates an annotated version tag named:

vYYYYMMDDHHMMSS-<commit-subject-slug>

It then pushes the current branch and tags to both configured remotes:

• github
• gitlink

Set AGENTBUS_AUTO_PUSH=0 to skip automatic pushing for one commit.