Skip to content

AgentFlow

Context engineering framework for multi-agent systems.

AgentFlow is a framework-agnostic toolkit for building multi-agent workflows where configuration lives in Markdown files with YAML front-matter. Define agents, workflows, routing rules, and memory -- all as readable .prompt.md, .workflow.md, .context.md, and .memory.md files.

Key Features

  • Markdown + YAML config files -- Define agents, workflows, and routing rules as human-readable Markdown with YAML front-matter. No JSON sprawl, no framework lock-in.

  • Pluggable LLM providers -- First-class support for Anthropic Claude, OpenAI GPT, Google Gemini, and any OpenAI-compatible API. Swap providers by changing one line of config.

  • Hybrid routing -- Combine deterministic YAML rules with LLM-based intent classification. Hierarchical domain routing groups agents under logical boundaries.

  • DAG-based workflows -- Compose agents into directed acyclic graphs with sync, parallel, and async execution modes. Pass outputs between nodes.

  • Sessions and memory -- Per-node scratchpads, session history, multi-user support, file-based memory, and vector search via Qdrant.

  • Tool registry -- Register tools with local Python handlers or HTTP endpoints. Agents call tools automatically during execution.

  • Event-driven observability -- Pub/sub event bus with built-in support for Langfuse telemetry. Hook into every framework event without coupling.

Quick Install

pip install gittielabs-agentflow

With LLM provider extras:

# Anthropic Claude
pip install "gittielabs-agentflow[anthropic]"

# OpenAI / OpenAI-compatible
pip install "gittielabs-agentflow[openai]"

# Google Gemini
pip install "gittielabs-agentflow[google]"

# Everything
pip install "gittielabs-agentflow[all]"

Minimal Example

1. Define an agent (context/agents/researcher.prompt.md):

---
name: researcher
provider: anthropic
model: claude-sonnet-4-6
temperature: 0.7
max_tokens: 4096
tools: [web_search, summarize]
---

You are a research agent. Given a topic, search for relevant information
and provide a comprehensive summary with sources.

2. Run it:

from agentflow import (
    ConfigLoader, AgentExecutor, ToolRegistry,
    SessionManager, EventBus, FileSystemStorage,
    AnthropicProvider,
)

# Load configs from the context/ directory
loader = ConfigLoader("./context")
loader.load()

# Set up infrastructure
storage = FileSystemStorage("./data")
events = EventBus()
provider = AnthropicProvider()
tools = ToolRegistry()
sessions = SessionManager(storage)

# Get agent config and execute
config, system_prompt = loader.get_agent("researcher")

Project Status

AgentFlow is in alpha (v0.5.0). The core APIs are stabilizing but may still change between minor versions. Production use should pin to a specific version.