Skip to content
Atlas by Comet Lab
Capability layer

AI agent frameworks, compared

An agent framework wires a model to tools and lets it plan, act and loop towards a goal. The choice turns on how much control flow you need to see, which language your team writes, and whether you are building on one model provider.

Learn the idea What an AI agent is, and what changes when AI takes action on Horizon

Open this layer in the explorer

At a glance

Every tool in this layer, with where it fits and the facts that usually decide it.

Tool Where it fits Licence Self-hostable Hosted API Maturity
LangGraph Graph orchestration Reach for LangGraph when an agent needs real branching (loops, retries, human-in-the-loop pauses) and you want that control flow written down explicitly. Open source Yes No Established
Agno Multi-agent framework A fit for Python teams building multi-agent systems where per-agent overhead and speed matter. Open source Yes No Growing
AutoGen Multi-agent conversation A fit for research-flavoured, exploratory multi-agent work where conversation between agents is the point. Open source Yes No Growing
Claude Agent SDK Lightweight agent SDK A strong fit when building agents on Claude that run long, use many tools, and need context handled rather than hand-managed. Open source Yes No Growing
CrewAI Multi-agent orchestration Good when a task decomposes cleanly into distinct roles and you want a fast, readable way to express that collaboration. Open source Yes Yes Growing
DSPy Prompt optimisation A fit when prompt quality is critical, you have examples to optimise against, and you want that tuning done by a compiler rather than by hand. Open source Yes No Growing
Mastra TypeScript agent framework The natural choice for TypeScript and Node teams who want agents in the same language and codebase as the rest of their app. Open source Yes No Growing
OpenAI Agents SDK Lightweight agent SDK A good fit when you want a minimal agent loop with handoffs and tracing, and you are already building on OpenAI models. Open source Yes No Growing
Pydantic AI Type-safe agent framework A strong fit for Python teams who want agent code that is statically typed and whose outputs are validated by construction. Open source Yes No Growing
Semantic Kernel Enterprise agent SDK The natural choice for .NET and enterprise Microsoft shops bringing agents into existing applications. Open source Yes No Established
smolagents Code-writing agents A fit for tasks where actions are naturally computational and a code-writing agent expresses them more cleanly than a chain of tool calls. Open source Yes No Growing

How to choose

Start from your situation. Where a line says “our”, it comes from Comet Lab’s own builds.

  1. The agent has real branching: loops, retries, human approval

    Control flow written down as an explicit graph, so a misfire becomes a named node you can inspect. Our default when an agent genuinely branches.

  2. The work splits cleanly into roles

    CrewAI expresses a crew of role-based agents quickly and readably; Agno suits Python multi-agent systems where per-agent overhead matters.

  3. Outputs feed other systems

    Typed agent code with outputs validated by construction: a malformed response fails loudly instead of poisoning the next step.

  4. You are building on one provider's models

    The Claude Agent SDK suits long-running agents on Claude, with context handled for you; the OpenAI Agents SDK is a minimal loop with handoffs and tracing for teams already on OpenAI models.

  5. Your team writes TypeScript or .NET

    Mastra keeps agents in the same TypeScript codebase as the app; Semantic Kernel brings agents into .NET and Microsoft-stack applications.

  6. You have examples to optimise against

    Prompt tuning done by a compiler instead of by hand. Without evaluation data, the ceremony outweighs the benefit.