pocket-joe 0.2.0.6

A correct, simple, performant, and pythonic framework for building durable AI agents

PocketJoe

LLM Agents are just agents...

• Agents are policies
• A policy reasons over observations and chooses a batch of options
• A policy can be any mix of LLM-based, human-in-the-loop, or heuristic

Semantics

An agent system using Reinforcement Learning theory with LLM semantics as first class

• policy: all code/logic/llm are policies
• observations - the set of observations for the policy to reason over
• options - additional action spaces available to the policy
• selected_actions - the set of concurrent actions the policy chose to take
• Message: a shared dataclass for observations and actions that aligns with llm semantics

LLM semantics as platform semantics

In LLM APIs, everything is a Message. We adopt this as our universal unit:

• Input: observations: list[Message] (what the policy sees)
• Output: selected_actions - the policy's action space (owns its outputs)

Key insight: When options are provided, they expand the policy's action space. The runtime automatically invokes all option calls and injects the results back as observations.

Everything is a Policy

Universal Return Types: Policies can return any JSON-serializable type - the framework automatically wraps results when called as options.

An LLM policy that can call other policies:

@policy.tool(description="Calls LLM with tool support")
async def openai_llm_policy_v1(observations: list[Message], options: list[OptionSchema]) -> list[Message]:
    """LLM policy that calls OpenAI GPT-4 with tool support."""
    openai = AsyncOpenAI()
    response = await openai.chat.completions.create(
        model="gpt-4",
        messages=observations_to_completions_messages(observations),
        tools=options_to_completions_tools(options))
    return completions_response_to_messages(response)

A simple helper policy returning primitives:

@policy.tool(description="Performs a web search and returns results.")
async def web_search_policy(query: str) -> str:
    """Performs a web search and returns results."""
    results = DDGS().text(query, max_results=5)
    return "\n\n".join([f"Title: {r['title']}\nURL: {r['href']}\nSnippet: {r['body']}" for r in results])

A policy returning structured data with Pydantic:

from pydantic import BaseModel

class TranscriptResult(BaseModel):
    title: str
    transcript: str
    thumbnail_url: str
    video_id: str
    error: str | None = None

@policy.tool(description="Transcribe YouTube video")
async def transcribe_youtube_policy(url: str) -> TranscriptResult:
    """Get video title, transcript and metadata from YouTube URL."""
    video_id = _extract_video_id(url)
    transcript = YouTubeTranscriptApi().fetch(video_id)
    return TranscriptResult(
        title=title,
        transcript=" ".join([snippet.text for snippet in transcript]),
        thumbnail_url=f"https://img.youtube.com/vi/{video_id}/maxresdefault.jpg",
        video_id=video_id
    )

An orchestrator policy that coordinates LLM + search:

@policy.tool(description="Orchestrates LLM with web search tool")
async def search_agent(prompt: str, max_iterations: int = 3) -> list[Message]:
    """Orchestrator that gives the LLM access to web search."""
    ctx = AppContext.get_ctx()

    system_builder = MessageBuilder(policy="system", role_hint_for_llm="system")
    system_builder.add_text("You are an AI assistant that can use tools.")
    system_message = system_builder.to_message()

    prompt_builder = MessageBuilder(policy="user", role_hint_for_llm="user")
    prompt_builder.add_text(prompt)
    prompt_message = prompt_builder.to_message()

    history = [system_message, prompt_message]
    for _ in range(max_iterations):
        selected_actions = await ctx.llm(
            observations=history,
            options=OptionSchema.from_func([ctx.web_search])
        )
        history.extend(selected_actions)
        if not any(msg.payload and isinstance(msg.payload, OptionCallPayload) for msg in selected_actions):
            break

    return history

Use AppContext for registry (gives IDE type hints):

class AppContext(BaseContext):
    def __init__(self, runner):
        super().__init__(runner)
        self.llm = self._bind(openai_llm_policy_v1)
        self.web_search = self._bind(web_seatch_ddgs_policy)
        self.search_agent = self._bind(search_agent)

Enjoy:

async def main():
    runner = InMemoryRunner()
    ctx = AppContext(runner)
    result = await ctx.search_agent(prompt="What is the latest Python version?")

    # Get final text message (Message.__str__ extracts text automatically)
    final_msg = next((msg for msg in reversed(result) if msg.parts), '')
    print(f"\nFinal Result: {final_msg}")

Why this matters:

• Universal Composability: Decorate any function - it works like FastAPI/FastMCP endpoints
• Flexible Return Types: Return primitives (str), Pydantic models, or list[Message] for complex flows
• Auto-wrapping: Framework automatically wraps results in OptionResultPayload when called as options
• Type-safe: Full IDE support with typed context and message payloads
• Evolution-friendly: Start simple (primitives) → add complexity (messages) with no refactoring

A correct, simple, performant, and pythonic framework for building durable AI agents.

"There is no flow, only Policies and Actions."

Getting Started

Prerequisites

• Python 3.12+

Installation

uv add pocket-joe

Or with pip:

pip install pocket-joe

To install with example dependencies:

uv add pocket-joe --extra examples
# or
pip install pocket-joe[examples]

Development Setup

git clone https://github.com/Sohojoe/pocket-joe.git
cd pocket-joe
uv sync --dev --all-extras

Running Examples

Set your API key:

export OPENAI_API_KEY=sk-...

Search Agent (ReAct)

uv run python examples/search_agent.py

YouTube Summarizer

uv run python examples/youtube_summarizer.py

Dev Status

Still in prerelease, things will change

Initial version

• [] Tidy up code - add partly refactored code
• [] Proper tests
• [] Implement more examples from Pocket-Flow

Durable System:

• [] Ledger - Temporal style 'at least once, only one result' replay semantic
• [] Durable Storage wrapper - For long running tasks & replay
• [] Distributed - worker model

Background

Inspired by PocketFlow... I loved PocketFlow but it fell short in a couple of key areas. This is my rewrite that I can actually use.