Python: [Feature]: Add a CodeAct backend for remote isolated Linux micro-VM sandboxes

[Patricio-Filice-Luxor]

Description

CodeAct currently has two backends, both in-process and local: agent-framework-hyperlight (WASM micro-VM) and agent-framework-monty (Rust Python interpreter). Both are excellent for fast, low-overhead code execution with host tool callbacks — but by design neither offers a real operating system: no apt/pip installs, no subprocesses or CLIs, no persistent filesystem, and the code ultimately executes on the host machine.

Agents that need a real computer — install packages, run shell commands, build projects, keep files across steps — currently have no CodeAct story.

Proposal: add agent-framework-tenki, wrapping Tenki Sandbox (managed Linux micro-VMs with a Python SDK) behind the same *CodeActProvider / *ExecuteCodeTool shape as the existing backends, so users can swap backends with minimal churn. Each sandbox is a remote, isolated, finite environment: code never touches the host, and server-side lifetime caps (max_duration, pause-snapshot retention) bound cost even if the client process crashes.

Why another backend next to Hyperlight/Monty:

	Hyperlight	Monty	Tenki
Execution	WASM micro-VM, in-process	Rust interpreter, in-process	remote managed Linux micro-VM
Code runs on host machine	yes (sandboxed)	yes (sandboxed)	no — fully remote
pip / apt installs	no	no	yes, persist across calls
Subprocesses / CLIs	no	no	yes
Real filesystem (persists across calls)	file mounts only	file mounts only	yes
Host tool callbacks	yes	yes	no (SDK has no bridge)
Platform limits	linux/x86_64, win/AMD64, py<3.14	cross-platform	any (remote service; needs network)

Additionally: seeding sandboxes from prepared snapshots. Snapshots and volumes are created and managed with the Tenki CLI/SDK, never by this package — e.g. provision a sandbox, install your toolchain, snapshot it. Agent sandboxes can then be seeded from that prepared state:

execute_code = TenkiExecuteCodeTool(
    extra_create_kwargs={"snapshot_id": "<your-snapshot-id>"},
)

The same passthrough works for persistent Tenki volumes, and with TenkiCodeActProvider — each run's fresh sandbox starts from the same prepared snapshot.

Note: Tenki is a managed service — using this backend requires a Tenki account and API key (TENKI_API_KEY), and sandbox usage is billed.

Code Sample

from agent_framework import Agent
from agent_framework_tenki import TenkiCodeActProvider

async with TenkiCodeActProvider() as codeact:
    agent = Agent(client=client, context_providers=[codeact])
    result = await agent.run(
        "Install polars, then compute the mean of column 'x' in /tmp/data.csv."
    )

Each agent run gets a fresh sandbox, terminated when the run completes.

Language/SDK

Python