harbor-multiagent

Harbor with several agents in one environment. Build any multi-agent simulation: games, a company of agents with coworkers and customers played by other agents, head-to-head competition, or self-play. Any Harbor harness and model can fill any role, under test or in support.

By Olam Labs. Not an official Harbor project.

Install

uv tool install git+https://github.com/olamlabs/harbor-multiagent

This installs the harbor command in place of an installed Harbor. Single-agent tasks run unchanged.

Example

Each agent runs in its own service of the task's environment/docker-compose.yaml, reads instructions/<service>.md, and is named in task.toml:

[agent]
timeout_sec = 600.0

[agents.p1]

[agents.p2]
under_test = false

A trial or job config puts a harness and model in each one:

participants:
  - seat: p1
    name: codex
    model_name: openai/gpt-5.5
  - seat: p2
    name: codex
    model_name: openai/gpt-5.5

Run the example task, examples/tasks/hello-seats, with Docker running and OPENAI_API_KEY set:

harbor trial start -c examples/configs/features/seats-trial.yaml
harbor run -c examples/configs/features/seats-job.yaml

docs-mintlify/tasks/multi-agent.mdx covers the task format, participants, how a run starts and ends, logs, the verifier, and the limitations.

Toward Harbor

We want multi-agent support in Harbor itself. Until then, this fork carries a working version on top of Harbor's main; CHANGES.md lists every Harbor file it touches and why. Issues and PRs are welcome.

License and citation

Apache 2.0, as Harbor. In academic work, please cite Harbor using CITATION.cff.