statek 1.1.3

Durable Python workspaces for stateful LLM agents

STATEK

STATEK (Stateful Agent Execution Kit) gives an LLM agent a durable Python workspace inside your application.

Instead of treating an agent as something outside the codebase that can only send JSON-shaped tool calls, STATEK lets the agent work with Python directly. A job can start with application objects already in scope, run model-written code against those objects, keep the variables it creates, pause, and resume later from the same execution state.

That makes it possible to embed agents into a Python system in a very literal way. The agent can use the same object model as the rest of the application, while STATEK keeps the job's local state, console output, chat history, tool calls, status, errors, and continuation state durable and inspectable.

One worker process can host many independent agent jobs, including large fleets of jobs that wait, resume, depend on one another, or run under bounded concurrency. In practical deployments this is meant to support anything from a few long-running workflows to tens of thousands of waiting or active jobs in a single process.

Read the full documentation at docs.dbzero.io/statek.

Where STATEK Fits

STATEK makes the most sense when an agent needs to work inside real application state, not just around it.

Use it for systems such as:

• data-intensive agents working in complex domains
• large-scale deployments where efficient orchestration matters
• human-in-the-loop workflows and long-running approval flows
• coordinated agent fleets with parent and child jobs
• workflows that route work across multiple LLM models or providers
• jobs that may wait for minutes, hours, or much longer before continuing

If your agent only needs to answer a single prompt or call a small set of stateless tools, STATEK may be more machinery than you need. It is most useful when the agent's work has memory, durable state, dependencies, or meaningful application-side effects.

Why Code-First Agents?

A code-first agent can choose what to do at each step:

1. answer or report something to the user
2. call a tool
3. execute a chunk of Python code

STATEK is built around the third option, while still supporting the first two. The important difference is that the Python execution is stateful and durable. Variables created in one step can still be used later, and a job can pause and resume without losing its working context.

That shape has some practical advantages:

• Python is often a better fit than plain JSON for expressing non-trivial work.
• Agents can use handles to real application objects instead of passing around large sets of opaque IDs.
• Durable, mutable state can be inspected and audited while the workflow is still in progress.
• LLMs are already strong Python generators, so it makes sense to let them use that skill directly.
• Agents, services, and application code can share one codebase and object model instead of meeting only through narrow integration layers.

See STATEK Core Concepts for the full mental model.

How It Works

A STATEK job is a persisted Python session for one unit of agent work. Application code, a dispatcher, or another agent can create a job with useful locals already available:

user
message
calendar
today
timestamp

Those names are not special framework APIs. They are just Python variables in the job's workspace. They can point to application objects, dbzero objects, service adapters, functions, or other controlled values your application chooses to expose.

The agent can then write ordinary Python against them:

events = calendar.events_for(today)
meeting = calendar.find_meeting("Weekly planning", day=today)
empty_slot = calendar.find_empty_slot(
    after=meeting.ends_at,
    duration=meeting.duration,
)
meeting.move_to(empty_slot)

STATEK persists the execution record that matters:

• Python locals
• console output
• chat history
• tool calls and results
• job status and continuation state
• errors and model usage where available
• references to durable application objects

Later, the same job can continue with the variables it created earlier. Other jobs can run in the same process without their local Python state colliding.

Quick Start

This is the minimal shape of a STATEK worker: install the package, open dbzero state, define an agent, start a worker, and submit work. The full walkthrough is in the STATEK Quickstart.

Install STATEK with dbzero support:

pip install "statek[dbzero]"

For deployments that need advanced dbzero-pro storage and security features, install:

pip install "statek[dbzero-pro]"

Initialize dbzero and open a prefix before constructing persisted STATEK objects:

import dbzero as db0
import statek

DBZERO_ROOT = "./statek-data"
PREFIX = "/acme/triage/dev/jobs"


def open_statek_store():
    db0.init(dbzero_root=DBZERO_ROOT, read_write=True)
    statek.init()
    statek.open_prefix(PREFIX, "rw")

Define a supervised agent:

from statek.agents.agent import SupervisedAgent
from statek.prompt_config import make_system_prompt


def create_triage_agent():
    agent = SupervisedAgent(
        role="triage_agent",
        _system_prompt=make_system_prompt(
            "You are a concise triage agent. Inspect the event and decide the next action."
        ),
        _tools=[],
        _metadata={"MODEL": "openrouter/openai/gpt-5-mini"},
    )

    agent.update_warmup_def(
        """
print("received event:", event)
"""
    )
    return agent

Start a blocking worker:

from statek import start_statek
from statek.statek_push_queue import StatekPushQueue


open_statek_store()

agent = create_triage_agent()
queue = StatekPushQueue(prefix=PREFIX)

start_statek(
    agents=[agent],
    push_queues=[queue],
    max_concurrency=10,
)

Submit work from a host service, script, webhook handler, or RPC endpoint:

from statek.statek_client_api import StatekClientAPI


open_statek_store()

agent = create_triage_agent()
incoming_event = {
    "type": "support_ticket",
    "subject": "Invoice total looks wrong",
    "priority": "normal",
}

job = StatekClientAPI().submit_new_job(
    agent=agent,
    shared_vars={"event": incoming_event},
    locale="EN-US",
    source="quickstart",
)

print("submitted job:", job)

dbzero and dbzero-pro

STATEK can run ordinary orchestration flows with dialogs and tool calls. The code-first model becomes much more powerful when application state lives in dbzero, because both your application and the agent can work with the same durable Python object graph.

With dbzero, durable application objects are normal Python classes decorated with @db0.memo:

import dbzero as db0


@db0.memo
class Meeting:
    def __init__(self, title, starts_at, ends_at):
        self.title = title
        self.starts_at = starts_at
        self.ends_at = ends_at

    def move_to(self, slot):
        self.starts_at = slot.starts_at
        self.ends_at = slot.ends_at

From the agent's point of view, this is still just Python:

meeting.move_to(empty_slot)

From the application's point of view, the mutation is durable state. The same object graph can be inspected by application code, workers, dashboards, and future agent steps.

Use dbzero-pro when deployments need advanced isolation features such as object-level security, protected fields, field masks, or data filtering policies.

Security Model

STATEK runs model-written Python in restricted mode by default. Restricted mode limits imports, builtins, reflection paths, hidden tool calls, source size, and AST size before code execution.

Keep restricted mode on for normal use:

import statek

statek.init()

Disable it only for trusted local development or when another isolation layer is responsible for the execution boundary:

import statek

statek.init(restricted=False)

Restricted mode is a useful defense layer, but it is not a complete production security model. The host application still owns:

• authorization and tenant boundaries
• secrets and provider credentials
• filesystem, network, CPU, memory, and process limits
• permissions around tools and application objects
• idempotency and policy for external side effects
• deployment monitoring and recovery procedures

Read STATEK Security & Sandboxing before exposing workers to real users or production data.

Documentation

STATEK guides:

• Overview
• Quickstart
• Core concepts
• Agents
• Jobs
• Tools
• Futures and temporal tools
• Model providers
• Operations
• Security
• API reference

dbzero:

• dbzero documentation
• dbzero-pro documentation

Development

Create and activate a virtual environment:

python -m venv venv
source venv/bin/activate

Install development dependencies:

pip install -e ".[dev]"

Run lint and tests:

./scripts/run_lint.sh
./scripts/run_tests.sh

Build the package:

./scripts/build.sh

License

Statek is licensed under the Apache License, Version 2.0. See LICENSE for the full license text and THIRD_PARTY_NOTICES.md for third-party dependency attributions.