agentixx 0.2.0

Sandboxed rollouts you call like typed Python (import as `agentix`)

Agentix

Sandboxed rollouts you call like typed Python.

Turn agents, tools, and scorers into Python callables. Package their dependencies into runtime images. Call them from evaluators, trainers, and orchestration code without writing a new runner for every pairing.

Documentation | Quickstart | Cookbook | Architecture

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.

You have	You expose	You call
Claude Code, Codex, Aider, OpenHands, or an internal agent	async def run(...) -> RunResult	await client.remote(run, ...)
Shell, files, repo setup, or local tools	async def run(command: str) -> BashResult	await client.remote(bash_run, ...)
SWE-bench, MLE-Bench, or an internal evaluator	async def score(...) -> Score	await client.remote(score, ...)

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