Blackgeorge: Python Agent Framework for LLM Tool-Calling and Multi-Agent Orchestration

A code-first Python framework for building AI agents, tool-calling workflows, and multi-agent systems with explicit APIs, structured outputs, safe tool execution, and pause/resume flows.

Works with OpenAI, Anthropic, DeepSeek, Gemini, Mistral, Ollama, and 100+ other providers through LiteLLM. Tools come from plain Python functions or MCP servers. Structured outputs are Pydantic models.

What you can build

• tool-calling AI agents with validated inputs
• multi-agent teams that coordinate work
• agentic workflows with parallel and sequential steps
• LLM services with durable run state, events, and resume

Core primitives

• Desk: orchestrates runs, events, and persistence
• Worker: single-agent execution with tools and memory
• Workforce: multi-worker coordination and management modes
• Workflow: step-based flows with parallel execution

Features

• tool execution with confirmation, user input, timeouts, retries, and cancellation
• structured output support with Pydantic models
• event streaming and run store persistence
• collaboration primitives: channel messaging and blackboard state
• memory stores including vector memory with configurable chunking
• LiteLLM adapter for OpenAI-compatible model providers
• MCP tool integration for external tool providers

Why Blackgeorge

Most teams start with a hand-written tool loop over the OpenAI or Anthropic SDK. It works until a tool needs human approval, a run has to survive a restart, or someone asks what a run cost. Blackgeorge is that loop with those parts built in: tool calls can pause for confirmation or user input, run state is stored and resumable from another process, every run reports token usage and cost with an optional budget, and structured outputs are validated with Pydantic and retried. The primitives stay small and explicit, so the execution flow reads like the code you would have written yourself.

How it compares

Cells describe each project's documented defaults; all four can be extended beyond them.

Use cases

• coding agents that edit files with confirmation and audit trails
• research and summarization agents with structured outputs
• support triage and routing across multiple workers
• operational workflows that pause for approvals and resume safely

See examples/coding_agent for a full end-to-end example.

Install

uv add blackgeorge

Vector memory is optional because ChromaDB adds a substantial dependency tree:

uv add "blackgeorge[vector]"

For development setup, see docs/development.md.

Quick start

A worker with one tool and a Pydantic output schema:

from pydantic import BaseModel

from blackgeorge import Desk, Job, Worker
from blackgeorge.tools import tool


class Summary(BaseModel):
    title: str
    bullets: list[str]


@tool()
def fetch_notes(topic: str) -> str:
    return f"Key points about {topic}: tool calling, structured output, pause and resume."


desk = Desk(model="openai/gpt-5-nano")
worker = Worker(name="Researcher", tools=[fetch_notes])
job = Job(input="Summarize agent frameworks using fetch_notes", response_schema=Summary)

report = desk.run(worker, job)
print(report.data.title, report.data.bullets)
print(report.metrics["cost_usd"])

Documentation

• Full documentation: jolovicdev.github.io/blackgeorge
• Searchable code walkthrough generated from the repository: DeepWiki
• Source for the docs lives in docs/; preview locally with uv run mkdocs serve

Job input

Job.input is the payload sent to the worker as the user message. If it is not a string, it is serialized to JSON. Use a string for simple requests, or a structured dict when you want explicit fields.

job = Job(
    input={
        "task": "Fix calculator behavior and update tests.",
        "context": "Use tools to inspect the project files.",
        "requirements": [
            "Confirm divide-by-zero behavior with the user.",
            "Confirm empty-average behavior with the user.",
            "Apply changes using tools.",
        ],
    },
    expected_output="Updated project files with consistent behavior.",
)

Workforce

from blackgeorge import Desk, Worker, Workforce, Job

desk = Desk(model="openai/gpt-5-nano")
w1 = Worker(name="Researcher")
w2 = Worker(name="Writer")
workforce = Workforce([w1, w2], mode="managed")

job = Job(input="Create a market report")
report = desk.run(workforce, job)

Workflow

from blackgeorge import Desk, Worker, Job
from blackgeorge.workflow import Step, Parallel

desk = Desk(model="openai/gpt-5-nano")
analyst = Worker(name="Analyst")
writer = Worker(name="Writer")

flow = desk.flow([
    Step(analyst),
    Parallel(Step(writer), Step(analyst)),
])

job = Job(input="Analyze product feedback")
report = flow.run(job)

Streaming

report = desk.run(worker, job, stream=True)

Pause and resume

from blackgeorge import Desk, Worker, Job
from blackgeorge.tools import tool

@tool(requires_confirmation=True)
def risky_action(action: str) -> str:
    return f"ran:{action}"

desk = Desk(model="openai/gpt-5-nano")
worker = Worker(name="Ops", tools=[risky_action])
job = Job(input="run risky")

report = desk.run(worker, job)
if report.status == "paused":
    report = desk.resume(report, True)

Session: multi-turn conversations

from blackgeorge import Desk, Worker

desk = Desk(model="openai/gpt-5-nano")
worker = Worker(name="ChatBot")

session = desk.session(worker)

session.run("My name is Alice")
session.run("What's my name?")

session_id = session.session_id

later_session = desk.session(worker, session_id=session_id)
later_session.run("Where do I live?")