Agent Framework Landscape

What Is an Agent Framework?

A library that handles the tool-calling loop: model generates a tool call → framework executes it → result fed back to model → repeat until done.

The hard parts they solve: - State management: Tracking conversation history, tool results, intermediate state - Tool orchestration: Defining, registering, and executing tools - Error recovery: Retrying, fallbacks, human-in-the-loop - Multi-agent coordination: Multiple agents collaborating or competing

The Landscape

OpenClaw

• What: Open-source Claude Code alternative
• Stack: TypeScript, pnpm, vitest
• Architecture: Single-agent tool loop with context pruning
• Strengths: Full Claude Code feature parity goal, CJK support, active development
• My experience: Contributed 4 PRs — stale tool detection, CJK token estimation, symlink bootstrap

LangGraph

• What: Stateful agent graphs from LangChain team
• Stack: Python, built on LangChain
• Architecture: Directed graphs with cycles, state machines for agent workflows
• Strengths: Persistence, human-in-the-loop checkpoints, complex multi-step workflows
• When to use: When your agent needs explicit state management and deterministic branching

CrewAI

• What: Multi-agent orchestration with roles
• Stack: Python
• Architecture: Role-based agents with defined goals, backstory, tools
• Strengths: Simple mental model (agents as team members), easy to set up
• Limitations: Less flexible than LangGraph for complex flows

AutoGen (Microsoft)

• What: Multi-agent conversation framework
• Stack: Python
• Architecture: Agents in group chats, conversational patterns
• Strengths: Group chat patterns, code execution, nested conversations
• Limitations: Complex API, heavy abstraction

smolagents (Hugging Face)

• What: Lightweight code agents
• Stack: Python
• Architecture: Agents write Python code directly (not JSON tool calls)
• Strengths: Simple, transparent, agents write real code
• Limitations: Less structured than tool-calling approaches

Pydantic AI

• What: Type-safe agent framework
• Stack: Python, Pydantic
• Architecture: Structured outputs + dependency injection
• Strengths: Type safety, Pydantic ecosystem, clean API
• When to use: When you want structured, validated outputs from agents

Comparison Matrix

Framework	Multi-Agent	State Mgmt	Tool Calling	Code Execution	Complexity
OpenClaw	No	Context window	JSON schema	Sandboxed bash	Medium
LangGraph	Yes	Graph state	LangChain tools	Via tools	High
CrewAI	Yes	Role-based	YAML/code	Via tools	Low
AutoGen	Yes	Chat history	Python	Built-in	High
smolagents	No	Minimal	Code-based	Direct Python	Low
Pydantic AI	No	DI container	Typed tools	Via tools	Medium

My Take

The framework wars are mostly about abstraction level. At the core, every framework does:

while not done:
    response = model.generate(messages, tools)
    if response.has_tool_call:
        result = execute_tool(response.tool_call)
        messages.append(result)
    else:
        done = True

The value is in the details: how they handle state, errors, concurrency, and multi-agent coordination. For simple single-agent use cases, you barely need a framework — a 50-line script works.