power-loop 0.8.0

Embeddable agent execution kernel — LLM loop, hooks, events, tools, dynamic sub-agents.

power-loop

Documentation | 中文文档 | Examples | Changelog

Embeddable, stateful agent execution for Python.

power-loop gives application code one small interface, StatefulAgentLoop, and handles the repetitive agent runtime work around it: multi-turn LLM loops, tool calls, hooks, events, context compaction, sub-agents, retry/cancel, structured output, memory, and SQLite-backed session persistence.

It is a library, not a service or a full application framework. You keep ownership of product logic, HTTP APIs, auth, queues, RAG, UI, and deployment.

Scope: orchestration, not isolation

power-loop orchestrates the agent loop; it does not sandbox tool execution. The built-in bash / file tools run in-process (a subprocess shell inheriting the host environment) — convenient for trusted, local use, but not a security boundary. If your agent runs model-authored or otherwise untrusted commands, run them in your own sandbox (container / gVisor / microVM) and inject it via the ShellBackend seam (runtime.exec_backend); power-loop launches the persistent shell through your backend. Keep secrets in your orchestrator — the loop does not scrub the tool environment for you.

Install

pip install power-loop

For local development:

git clone https://github.com/PL-play/power-loop.git
cd power-loop
pip install -e ".[dev]"

Python 3.10+ is required.

Quick Example

import asyncio

from power_loop import AgentLoopConfig, StatefulAgentLoop, create_llm_service_from_env


async def main() -> None:
    llm = create_llm_service_from_env()
    loop = StatefulAgentLoop(
        llm=llm,
        db_path="./power_loop_sessions.db",
        config=AgentLoopConfig(
            system_prompt="You are a concise assistant.",
            max_rounds=4,
        ),
    )

    sid = loop.new_session(metadata={"user_id": "demo"})
    first = await loop.send("My favorite color is teal.", session_id=sid)
    second = await loop.send("What is my favorite color?", session_id=sid)

    print(second.final_text)


asyncio.run(main())

Configure any OpenAI-compatible endpoint with environment variables:

POWER_LOOP_BASE_URL=https://api.openai.com/v1
POWER_LOOP_API_KEY=sk-...
POWER_LOOP_MODEL=gpt-4o-mini

See Getting Started for the complete first run.

What It Provides

Capability	Where to read more
Stateful sessions and cross-process resume	Sessions
Tool calling with JSON Schema validation	Tools
Lifecycle hooks for control flow	Hooks
Typed events for streaming, audit, and metrics	Events
Context compaction	Compaction
Sub-agents with AgentSpec	Sub-agents
Retry, timeout, and cancellation	Retry & Cancel
Structured JSON output	Structured Output
Pluggable cross-session memory	Memory
Provider configuration	Providers

Per-call overrides

Build one loop and reuse it across callers; restrict tools or swap the system prompt per send without rebuilding (the model only sees the allowed tools). Ideal for multi-tenant hosts.

# loop registered with all tools; this run exposes only "get_weather"
await loop.send("…", session_id=sid, tools=["get_weather"])

# per-run system prompt override (precedence: per-call > session > config)
await loop.send("…", session_id=sid, system_prompt="You are a terse bot.")

The same overrides are available on send_sync(). When follow_up() is idle and falls back to a new send, it accepts them too. A follow-up queued into an already running call keeps that call's active tool and prompt policy.

For a multi-tenant host that reuses one registry across workspaces, build an unbound registry and supply the workspace at invocation time:

from power_loop import RuntimeEnv, create_default_tool_registry, runtime_env_context

registry = create_default_tool_registry(preset="core", bind=False)
with runtime_env_context(RuntimeEnv(workspace_dir=tenant_workspace)):
    result = await registry.invoke_async("read_file", {"path": "README.md"})

See examples/23_per_send_overrides.py.

Public API

Stable imports are re-exported from power_loop:

from power_loop import (
    AgentLoopConfig,
    StatefulAgentLoop,
    StatefulResult,
    ToolDefinition,
    ToolRegistry,
)

The stability tiers are:

Tier	Meaning
Stable	Backward compatible across minor releases. Listed in power_loop.STABLE_API.
Provisional	Available from the top-level package during 0.x, but may change.
Internal	Submodule imports such as power_loop.core.*; no compatibility promise.

See the API reference for the current surface.

Examples

The examples/ directory is ordered from minimal usage to full chatbot composition:

python examples/00_hello_world.py
python examples/02_tool_calling.py
python examples/19_full_chatbot.py

The full list is in examples/README.md.

Development

pip install -e ".[dev]"
ruff check .
pytest -q --no-real

Real LLM examples/tests use POWER_LOOP_* or the legacy OPENAI_COMPAT_* variables.

Project Links

• Documentation index
• Architecture
• Roadmap
• Changelog
• Contributing
• License