graphbus 0.5.1

Multi-agent orchestration framework — LLM-powered agents negotiate, refactor, and evolve your codebase

GraphBus

Distribute your AI reasoning. Multiple specialized agents negotiate, refactor, and evolve your codebase — separation of concerns at the reasoning level.

Website · Quickstart · Examples · CLI Reference · Architecture

---

What is GraphBus?

GraphBus is a Python framework with a radical idea: let your agents improve the code itself, not just run it.

Every class in a GraphBus project is a potential LLM agent. During a build cycle, agents wake up, read their own source, propose improvements, and negotiate consensus with other agents via a typed message bus. An arbiter resolves conflicts. The result is committed back to source.

At runtime, agents execute their negotiated logic. Each agent owns a focused slice of responsibility — no single context window trying to reason about everything at once.

Negotiate (agents improve the codebase) → Deploy (agents run their domain logic)

Why this matters

Most LLM frameworks use a single context window that grows unbounded — Claude Code, Cursor, Copilot. GraphBus inverts this: specialized agents each reason about their own domain, negotiate at boundaries, and commit improved code. The cognitive load is distributed, not bottlenecked.

---

Getting Started

GraphBus uses two keys for different purposes:

LLM provider key — powers LLM agent negotiation via LiteLLM. Set the key for your chosen provider:

export DEEPSEEK_API_KEY=...           # default model: deepseek/deepseek-reasoner
export ANTHROPIC_API_KEY=sk-ant-...   # for claude-* models
export OPENAI_API_KEY=sk-...          # for gpt-* models
export OPENROUTER_API_KEY=...         # access all models with one key

GraphBus API key (optional) — warehouses your negotiation history, contracts, and cross-session memory at api.graphbus.com. Without it, negotiation works fine but history isn't persisted.

export GRAPHBUS_API_KEY=gb_...        # sign up at graphbus.com

---

Quickstart

# Install
pip install graphbus

# Create a new project
graphbus init my-project --template microservices
cd my-project

# Build the agent graph
graphbus build agents/

# Tell your agents what to do
graphbus negotiate .graphbus --intent "add retry logic to all API calls"

The --intent flag is the primary way to use GraphBus. You describe what you want changed in plain English, and your agents negotiate how to implement it — proposing diffs, evaluating each other's work, and committing consensus changes.

# More intent examples
graphbus negotiate .graphbus --intent "add input validation to all endpoints"
graphbus negotiate .graphbus --intent "refactor error handling to use custom exceptions"
graphbus negotiate .graphbus --intent "add rate limiting to the payment service"

# Scope to a namespace
graphbus negotiate .graphbus -n backend-api --intent "reduce latency in hot paths"

# Run the built artifacts (zero AI cost at runtime)
graphbus run .graphbus/

---

Hello World

# agents/hello_service.py
from graphbus_core import GraphBusNode, schema_method, subscribe

class HelloService(GraphBusNode):
    SYSTEM_PROMPT = "I generate friendly greeting messages."

    @schema_method(
        input_schema={},
        output_schema={"message": str}
    )
    def generate_message(self):
        return {"message": "Hello from GraphBus!"}

    @subscribe("/Hello/MessageGenerated")
    def on_message(self, event):
        self.log(event.payload)

graphbus build agents/
# [BUILD] Scanning agents/hello_service.py
# [BUILD] Graph: 1 node, 0 edges
# [BUILD] Artifacts written to .graphbus/

graphbus run .graphbus/
# [RUNTIME] Loaded 1 agent
# [RUNTIME] HelloService → "Hello from GraphBus!"

Enable agents and watch them negotiate:

export DEEPSEEK_API_KEY=your_key_here  # or ANTHROPIC_API_KEY=sk-ant-...
graphbus build agents/ --enable-agents
# [AGENT] HelloService: "I propose adding input validation..."
# [AGENT] LoggerService: "I accept — improves contract safety"
# [ARBITER] Consensus reached. Committing changes.
# [BUILD] Artifacts written to .graphbus/ (2 files modified)

---

Architecture

GraphBus has two strictly separated modes:

┌─────────────────────────────────────────────────────────────────┐
│                        BUILD MODE                               │
│  ┌──────────┐  proposals  ┌─────────┐  evaluations  ┌────────┐ │
│  │ AgentA   │────────────▶│  BUS    │◀──────────────│ AgentB │ │
│  │  (LLM)   │◀────────────│         │───────────────▶│  (LLM) │ │
│  └──────────┘  commits    └────┬────┘               └────────┘ │
│                                │                                │
│                          ┌─────▼─────┐                         │
│                          │  Arbiter  │  resolves conflicts      │
│                          └─────┬─────┘                         │
│                                │                                │
│                    ┌───────────▼──────────┐                     │
│                    │   Build Artifacts     │  (.graphbus/)       │
│                    │  graph.json           │                     │
│                    │  agents.json          │                     │
│                    │  topics.json          │                     │
│                    └───────────────────────┘                     │
└─────────────────────────────────────────────────────────────────┘
                              │
                              ▼ (deploy once)
┌─────────────────────────────────────────────────────────────────┐
│                       RUNTIME MODE                              │
│                                                                 │
│  ┌──────────┐   events    ┌─────────┐   events    ┌──────────┐ │
│  │ AgentA   │────────────▶│  BUS    │────────────▶│ AgentB   │ │
│  │ (static) │             │ (pub/   │             │ (static) │ │
│  └──────────┘             │  sub)   │             └──────────┘ │
│                            └─────────┘                          │
│                                                                 │
│          ✅ No LLM calls   ✅ Deterministic   ✅ $0 AI cost     │
└─────────────────────────────────────────────────────────────────┘

Core Concepts

Concept	Description
GraphBusNode	Base class for all agents. Subclass it, add a SYSTEM_PROMPT, decorate methods.
@schema_method	Declares typed input/output schema for a method — forms the contract between agents.
@subscribe	Registers a handler for a topic on the message bus.
@depends_on	Declares a dependency edge between agents in the DAG.
Build Artifacts	JSON files emitted after a build: graph.json, agents.json, topics.json.
Arbiter	A special agent (mark with IS_ARBITER = True) that resolves conflicting proposals.
Message Bus	Typed pub/sub backbone. Topics are typed paths (e.g. /Order/Created).

---

CLI Reference

GraphBus ships a full-featured CLI with 18 commands:

graphbus [OPTIONS] COMMAND [ARGS]...

Core Commands

Command	Description
graphbus build <path>	Scan agents, build dependency graph, emit artifacts
graphbus run <artifacts>	Load artifacts and execute the runtime
graphbus inspect <artifacts>	Inspect build artifacts (graph, agents, topics)
graphbus validate <path>	Validate agent definitions without building
graphbus tui	Launch interactive TUI (keyboard-driven UI)

Development Tools

Command	Description
graphbus init <name>	Initialize a new project from template
graphbus generate agent <Name>	Generate agent boilerplate
graphbus profile <artifacts>	Profile runtime performance
graphbus dashboard	Launch web-based visualization dashboard
graphbus negotiate <path> --intent "..."	Main command — tell agents what to improve
graphbus inspect-negotiation	Browse negotiation history

Deployment Tools

Command	Description
graphbus docker build	Generate Dockerfile for your project
graphbus docker run	Build and run in Docker
graphbus k8s generate	Generate Kubernetes manifests
graphbus k8s deploy	Deploy to Kubernetes cluster
graphbus ci github	Generate GitHub Actions workflow
graphbus ci gitlab	Generate GitLab CI pipeline

Advanced

Command	Description
graphbus state	Manage agent state persistence
graphbus coherence	Run inter-agent coherence checks
graphbus contract	Validate schema contracts between agents
graphbus migrate	Migrate artifacts across schema versions

---

Examples

Three working examples are included in examples/:

1. hello_graphbus — The basics

cd examples/hello_graphbus
python build.py              # Build without agents
DEEPSEEK_API_KEY=your_key python build.py   # Build with LLM agents
python run.py                # Run the built artifacts

2. hello_world_mcp — MCP integration

GraphBus ships an MCP (Model Context Protocol) server so any MCP-compatible client can interact with a running GraphBus runtime as a tool.

cd examples/hello_world_mcp
graphbus build agents/
graphbus run .graphbus/ --mcp   # Exposes MCP endpoint

3. news_summarizer — Real-world pipeline

A multi-agent news summarization pipeline. One agent fetches, one summarizes, one formats. Each agent owns its domain. Agents negotiate a shared schema for the summary output and run their specialized logic at runtime.

cd examples/news_summarizer
graphbus build agents/
OPENAI_API_KEY=sk-... graphbus run .graphbus/

---

The Negotiation Protocol

When --enable-agents is set, each agent gets an LLM instance. Build Mode runs this cycle:

1. SCAN     → Discover all GraphBusNode subclasses in the target path
2. EXTRACT  → Parse methods, schemas, subscriptions, system prompts
3. BUILD    → Construct networkx DAG (topological sort for eval order)
4. ACTIVATE → Instantiate one LLM agent per node
5. PROPOSE  → Each agent reads its source and proposes improvements
6. EVALUATE → Agents evaluate each other's proposals (accept/reject + reasoning)
7. ARBITRATE → Arbiter resolves split decisions
8. COMMIT   → Accepted proposals are applied to source files
9. ARTIFACT → Build graph + agent metadata serialized to .graphbus/

Proposals are structured messages:

class Proposal:
    agent_id: str
    target_file: str
    diff: str          # unified diff
    rationale: str     # LLM reasoning
    affects: list[str] # other agents impacted

---

Project Structure

graphbus-core/
├── graphbus_core/           # Core library
│   ├── node_base.py         # GraphBusNode base class
│   ├── decorators.py        # @schema_method, @subscribe, @depends_on
│   ├── config.py            # BuildConfig, RuntimeConfig
│   ├── build/               # Build pipeline (scanner, extractor, builder, writer)
│   ├── runtime/             # Runtime engine (loader, bus, router, executor)
│   ├── agents/              # LLM agent wrappers
│   └── model/               # Pydantic models (Message, Event, Proposal, ...)
├── graphbus_cli/            # CLI (click + rich)
│   ├── main.py              # Entry point
│   ├── commands/            # One file per command group
│   └── repl/                # Interactive REPL
├── graphbus_api/            # REST API server
├── graphbus-mcp-server/     # MCP protocol server
├── examples/
│   ├── hello_graphbus/      # Basic example
│   ├── hello_world_mcp/     # MCP integration
│   └── news_summarizer/     # Real-world pipeline
├── tests/                   # Full test suite
└── docs/
    └── core/                # Architecture docs

---

vs. LangGraph / CrewAI / AutoGen

	GraphBus	LangGraph	CrewAI	AutoGen
Agents rewrite source code	✅ Core feature	❌	❌	⚠️ Limited
Zero LLM cost at runtime	✅ Always	❌ Every call	❌ Every call	❌ Every call
Agent negotiation / consensus	✅ Built-in	❌	⚠️ Partial	⚠️ Partial
Graph-native DAG orchestration	✅ networkx	✅	❌	❌
Typed schema contracts per edge	✅	⚠️ Partial	❌	❌
Build / Runtime mode separation	✅ Core design	❌	❌	❌
Full deployment tooling (K8s/Docker)	✅ CLI native	❌	❌	❌
Interactive TUI	✅	❌	❌	❌

The key difference: other frameworks run agents to perform tasks. GraphBus runs agents to improve the code that performs tasks. After a build cycle, the intelligence is baked into static artifacts — not perpetually consumed at runtime.

---

Installation

From source (current)

git clone https://github.com/graphbus/graphbus-core
cd graphbus-core
pip install -e .

From PyPI

pip install graphbus

# With TUI support
pip install graphbus[tui]

Requirements

• Python 3.9+
• networkx >= 3.0
• click >= 8.1.0
• rich >= 13.0.0

Optional (for LLM agents):

• litellm — all LLM providers (Anthropic, OpenAI, DeepSeek, OpenRouter, etc.)

Optional (for TUI):

• textual >= 0.47.0

---

Testing

# Run all tests
pytest

# With coverage
pytest --cov=graphbus_core --cov-report=term-missing

# Run a specific test suite
pytest tests/test_runtime/
pytest tests/test_build/

Test coverage:

• Build pipeline: 100% passing (scanner, extractor, graph builder, artifact writer)
• Runtime engine: 100% passing (loader, message bus, event router, executor)
• End-to-end: Hello World example builds and runs clean
• CLI: All commands smoke-tested

---

Contributing

GraphBus is in alpha and we welcome contributors. See CONTRIBUTING.md for guidelines.

Quick start for contributors:

git clone https://github.com/graphbus/graphbus-core
cd graphbus-core
pip install -e ".[dev]"
pytest                    # Make sure everything passes

Areas where we especially want help:

• More LLM backends — LiteLLM integration supports many providers; help us test them
• More examples — real-world pipelines showing agent negotiation
• Documentation — architecture docs, tutorials, protocol spec
• Benchmarks — latency/cost comparisons vs. runtime LLM frameworks

---

Roadmap

See ROADMAP.md for the full roadmap with targets and status.

What's shipped (v0.1 alpha):

• Build Mode (scanner → extractor → graph builder → artifact writer)
• Runtime Mode (loader → message bus → event router → executor)
• CLI with 18 commands
• LLM negotiation engine (propose / evaluate / arbitrate / commit)
• MCP server integration
• Docker + Kubernetes deployment tooling
• 800+ tests, CI with GitHub Actions

Coming next (v0.2):

• graphbus dev — hot-reload mode during development
• Message trace UI — replay message flows in a web UI
• graphbus test — agent unit tests with full runtime wired in

Later:

• PyPI release (pip install graphbus)
• Multi-provider LLM support (via LiteLLM)
• Ollama local LLM backend
• Multi-process distributed runtime
• TypeScript SDK
• Protocol specification (for non-Python implementations)

Want to influence what ships next? Open a GitHub Discussion or 👍 the relevant issue.

---

License

MIT. See LICENSE.

---

Links

• 🌐 graphbus.com — Landing page + waitlist
• 📧 [e-mail] — Questions, feedback, partnership