The 10-Second Model

Agentix has two primitives:

• Remote calls: client.remote(fn, *args, **kwargs) runs a Python callable inside a sandbox worker. The callable is serialized with stdlib pickle, Python's native callable reference mechanism.
• Bundles: agentix build [path] packages a Python project and its declared dependencies into a deploy-ready runtime image.

from agentix import RuntimeClient, SandboxConfig, session
from agentix.bash import run
from agentix.deployment.docker import DockerDeployment

config = SandboxConfig(
    image="python:3.13-slim",
    runtime_image="hello-agentix:0.1.0",
)
async with session(DockerDeployment(), config) as sandbox:
    async with RuntimeClient(sandbox.runtime_url) as client:
        result = await client.remote(run, command="echo hello from $(uname -a)")

The unit of composition is not a bespoke benchmark runner or agent adapter. It is a Python callable.

Why Agentix Exists

Agent experiments sprawl quickly. One agent needs a CLI wrapper. Another needs a Python harness. A benchmark needs repo setup, grading scripts, and logs. A training loop needs the same pieces batched across many sandboxes.

Agentix collapses that matrix into one execution contract: if Python can serialize the callable and the sandbox has its dependencies, the host can call it.

What Ships

• Unary remote calls (await client.remote(fn, *args, **kwargs)) across the host-to-sandbox boundary. One shape. Pickled callable + pickled args/kwargs go in; the pickled return value comes back.
• One runtime worker process today behind an internal worker backend boundary, so future pools or per-call isolation can stay API-compatible.
• Tracing as an orthogonal layer: agentix.trace.span(...) inside the sandbox produces Trace/Span/SpanEvent lifecycle events that cross to the host via a side-channel and dispatch to any agentix.trace.Processor the host has registered.
• Bundle builds from normal Python projects and pyproject.toml dependencies.
• Optional Nix system dependencies when a project includes default.nix.
• Deployment backend plugins through the agentix.deployment entry point group.

Quickstart

Run the smallest demo from agentix-cookbook/examples/hello-agentix:

cd examples/hello-agentix
uv sync
uv run agentix build . --name hello-agentix  # builds hello-agentix:0.1.0
uv run python run.py

The demo builds hello-agentix:0.1.0, overlays it onto python:3.13-slim, then calls agentix.bash.run inside the sandbox:

import asyncio

from agentix import RuntimeClient, SandboxConfig, session
from agentix.bash import run
from agentix.deployment.docker import DockerDeployment


async def main() -> None:
    deployment = DockerDeployment()
    config = SandboxConfig(
        image="python:3.13-slim",
        runtime_image="hello-agentix:0.1.0",
    )
    async with session(deployment, config) as sandbox:
        print(f"sandbox up at {sandbox.runtime_url}")
        async with RuntimeClient(sandbox.runtime_url) as client:
            result = await client.remote(run, command="echo hello from $(uname -a)")
            print(f"exit={result.exit_code} stdout={result.stdout!r}")


asyncio.run(main())

Read the full quickstart for the project layout, lockfile, and runtime-image details.

Architecture

Host process
  RuntimeClient.remote(fn, *args, **kwargs)
    builds RemoteCallable from pickle.dumps(fn)
    pickles (args, kwargs) as one blob
        |
        v  Socket.IO "unary"
Sandbox
  agentix-server
        |  msgpack frame
        v
  worker subprocess
    RemoteCallable.resolve() -> fn
    pickle.loads(arguments) -> args, kwargs
    calls fn(*args, **kwargs)
    pickles the result back

The wire carries a single shape: unary. Cancellation has its own event (cancel). Trace lifecycle events cross as a separate broadcast channel (trace:event). HTTP is kept only for /health. Errors stay in-band.

Repository Map

• Agentix-Runtime-Basic: sandbox primitives such as bash and file operations.
• Agentix-Deployment-Docker: local Docker deployment backend.
• Agentix-Deployment-Daytona and Agentix-Deployment-E2B: hosted sandbox backend packages.
• agentix-cookbook: working integration recipes for agents and benchmarks.
• abridge: rollout-to-RL-buffer bridge.

Development

git clone https://github.com/Agentiix/Agentix
cd Agentix
pip install -e '.[dev]'
pytest
ruff check agentix/ tests/

Pair this repo with sibling backend/runtime repos checked out next to it when testing full sandbox rollouts.

Links

• Docs
• Quickstart
• Remote calls
• Bundles
• Architecture
• Roadmap