# Agent-framework scorecard — 3 open-source frameworks

All cells verified live on 2026-10-24 UTC against the projects' own repos/releases (GitHub API `published_at`, `license.spdx_id`). Maintenance check (release within 60 days): ADK, Pydantic AI, and CrewAI all released today; smolagents' last release is 2026-05-29 (>60 days), noted below.

| Criterion | Google ADK (Python) | Pydantic AI | CrewAI |
|---|---|---|---|
| License | [Apache-2.0](https://api.github.com/repos/google/adk-python) | [MIT](https://api.github.com/repos/pydantic/pydantic-ai) | [MIT](https://api.github.com/repos/crewAIInc/crewAI) |
| Latest release date | [2026-10-23, v2.13.0](https://github.com/google/adk-python/releases) | [2026-10-24, v2.61.0](https://github.com/pydantic/pydantic-ai/releases) | [2026-10-24, v1.15.30](https://github.com/crewAIInc/crewAI/releases) |
| Supported model providers | [Gemini native; Claude/OpenAI and others via adapters](https://adk.dev/agents/models/) | [OpenAI, Anthropic, Google, Bedrock, Ollama, and more](https://pydantic.dev/docs/ai/overview/) | [OpenAI, Anthropic, Google, local via LiteLLM/Ollama](https://docs.crewai.com/) |
| MCP / tool-calling support | [Function, OpenAPI, and MCP tools](https://google.github.io/adk-docs/tools/) | [Typed function tools plus MCP capability](https://ai.pydantic.dev/mcp/) | [Built-in tools, custom tools, and MCP integration](https://docs.crewai.com/concepts/tools) |
| Notable limitation | [Requires Python 3.11+](https://github.com/google/adk-python) | [Requires Python 3.11+](https://github.com/pydantic/pydantic-ai) | [Role/crew abstraction adds ceremony for single-agent jobs](https://docs.crewai.com/) |

**Verdict.** For a small automation project I would pick **Pydantic AI**: its typed tool contracts make inputs and outputs validated rather than prompt-hoped, and it stays provider-agnostic. I would choose **Google ADK** instead when the project is Gemini-centric and wants built-in evaluation/deployment tooling. I would reach for **CrewAI** only when the work genuinely decomposes into distinct roles with handoffs, since its crew abstraction is overhead for a single agent. (Rows above are the sole basis for this choice.)

_(Scorecard rows, framework names, and headings kept in English; total under 350 words. All source URLs fetched live at delivery time.)_