Deep Agents on Harakiri

An optional backend for the real Python deepagents framework. Developer preview 0.1.0rc1 targets Deep Agents 0.7.21 and Python SDK 0.1.0rc1 exactly, on Python 3.11 or newer. See the release record for publication status, exact artifacts and native qualification evidence.

Version-pinned PyPI installation, including the exact SDK and framework dependencies:

python -m pip install "h-sandbox-deepagents==0.1.0rc1"

from deepagents import create_deep_agent
from harakiri import HarakiriClient
from harakiri_deepagents import HarakiriSandboxBackend

with HarakiriClient.from_env() as client:
    with client.sandboxes.task(template="your-installed-template") as sandbox:
        backend = HarakiriSandboxBackend(sandbox)
        agent = create_deep_agent(model=model, backend=backend)
        result = agent.invoke({"messages": [{"role": "user", "content": prompt}]})

Configure model and prompt in your application; complete examples live in examples/python-first-task. The model and agent loop stay in the application. The sandbox runs shell and file tools. Neither model nor API credentials are forwarded into it. A backend always borrows the supplied sandbox.

Use AsyncHarakiriSandboxBackend with AsyncHarakiriClient and agent.ainvoke() for native async I/O. Use a borrowed sandbox, not a disposable task context, when the graph pauses for human approval. Checkpoints do not extend runtime TTL.

The complete integration guide covers model setup, installation, template utilities, partial transfers, truncation, observed command recovery and current qualification status. Standard output is bounded to 64 KiB by default; abnormal termination remains visible to the actual framework tools. No distributed exactly-once recovery is claimed.