h5i-python: Programmable Multi-Agent Orchestration

Claude Code, Codex, and other coding agents have different strengths. However, naive multi-agent orchestration such as simply launching several agents in parallel or allowing them to exchange messages does not define a reproducible development process. A real workflow must specify:

• who implements;
• who reviews whom;
• when an agent must revise its work;
• which candidates are independently tested;
• how the winner is selected; and
• when the selected change is applied to the original branch.

h5i-python is the Python SDK for the h5i orchestra engine. This SDK lets you define and execute multi-agent coding workflows across Claude Code, Codex, and other runtimes as ordinary Python programs.

Each agent works inside its own sandboxed Git worktree, so it cannot overwrite the original checkout or another agent's work. Agent turns produce Git-backed artifacts that can be reviewed, revised, neutrally verified, compared, selected, and applied as one auditable workflow.

1. Install

Install the h5i engine:

curl -fsSL https://raw.githubusercontent.com/h5i-dev/h5i/main/install.sh | sh
# cargo install --git https://github.com/h5i-dev/h5i h5i-core

Install the Python SDK from GitHub:

pip install h5i-orchestra
# pip install "git+https://github.com/h5i-dev/h5i-python.git"

2. Quickstart

Create ensemble.py inside the Git repository the agents should modify. This workflow let Claude and Codex independently implement the same task, review and improve each other’s work, and then select the better result.

from h5i.orchestra import Conductor

async def main(task):
    async with Conductor(repo=".", run="demo-task", launcher="resident") as c:
        claude = await c.hire("claude-agent", runtime="claude")
        codex  = await c.hire("codex-agent",  runtime="codex")

        # Have both agents implement the task independently and in parallel
        claude_work, codex_work = await asyncio.gather(claude.work(task), codex.work(task))

        await c.freeze() # Seal the round, ensuring that neither agent influenced the other beforehand

        # Have each agent review the other's work
        await asyncio.gather(codex.review(claude_work), claude.review(codex_work))

        # Verify each submission in a fresh, neutral sandbox
        await c.verify(claude_work, ["pytest", "--quiet"])
        await c.verify(codex_work, ["pytest", "--quiet"])

        verdict = await c.judge() # Select the smallest diff among the submissions that pass all tests
        print("winner:", verdict.selected_submission)

asyncio.run(main("implement quicksort in python with unit test"))

Run it as a normal Python program:

python ensemble.py

With the default launcher="resident", h5i automatically starts the agent sessions through tmux.

If you use herdr (an agent multiplexer for the terminal), pass launcher="herdr" instead: each agent seat comes up as a herdr pane beside your work, labeled h5i-orch-<run>-<agent>, with herdr's own per-agent status (working / blocked / done) in its sidebar.

3. Examples

examples/tutorial provides basic multi-agent orchestration patterns:

• arena_score.py: independent arena ranking;
• ensemble_score.py: mutual-review ensembles;
• debate_then_build.py: architect-to-implementer pipelines;
• review_escalation.py: conditional review escalation;
• judge_panel_score.py: LLM judge panels;
• tournament.py: tournament brackets; and
• custom_control_flow.py: custom Python control flow.

examples/papers/ re-implements the core workflow of 40 published multi-agent papers:

4. License

Apache-2.0