Pick one: a decision, not a bake-off

Which framework — if any?

Interactive decision tree — outcomes:

• No framework — write the loop

  Own the ~100 lines end to end: full auditability, zero abstraction debt, and the education you need before you can judge any framework. Revisit when you find yourself building checkpointing and a trace viewer by hand.

• Vercel AI SDK

  The TypeScript toolkit for model calls and streaming UI across Next.js, React, Svelte, Vue, Angular, and Node — with agent abstractions layered on top (v7, Apache-2.0, as of 2026-09). If the agent outgrows the UI layer, pair it with or graduate to Mastra.

• Mastra

  The full-featured TypeScript agent backend: agents, graph workflows, suspend/resume human-in-the-loop, RAG, evals, MCP (Apache-2.0 core, source-available ee/). Young company (YC W25) — weigh the velocity against the track record.

• OpenAI Agents SDK

  Lightweight agents, handoffs, guardrails, sessions, tracing — and provider-agnostic despite the name. Caveat that matters: pre-1.0 (v0.22.x as of 2026-09), so pin versions and read changelogs.

• Claude Agent SDK

  You inherit the production Claude Code harness — gather context → act → verify → repeat, subagents with isolated contexts, automatic compaction — instead of assembling one. Eyes open: this is the deepest provider coupling on the page (the package bundles the Claude Code CLI; usage under Anthropic’s commercial terms), and it is pre-1.0.

• LangGraph (or typed durable execution)

  Explicit graph, checkpointed state you can persist, inspect, and replay — the strongest audit story, stable at 1.x since October 2025. If your team is Pydantic-native and the need is durability more than graph audit, Pydantic AI’s Temporal/DBOS/Prefect/Restate integrations are the typed alternative.

• Strands Agents

  AWS’s model-driven, in-process loop — no hosted control plane, OpenTelemetry by default, proven inside Amazon Q Developer and AWS Glue before it was open-sourced (Apache-2.0). Deploys naturally to Bedrock AgentCore, which also hosts LangGraph and CrewAI agents if you change your mind.

• Microsoft Agent Framework

  The declared successor to Semantic Kernel and AutoGen, production-ready at 1.0+ with an LTS commitment, in Python and .NET (Go in preview). The default answer on an Azure estate — and the only first-class .NET answer on this page.

• Google ADK

  Code-first, Apache-2.0, five languages, the paved road into Google Cloud’s managed agent runtime — and the framework Google runs its own agent products on. Note the 2.x line broke from 1.x; check migration notes before upgrading.

• CrewAI

  Role-based Crews for autonomous collaboration, Flows when you need event-driven control back (MIT, 1.15.x as of 2026-09, standalone — no LangChain underneath). The fastest path from a team-of-specialists idea to running code; add Flows early if auditability matters.

• Pydantic AI

  Typed agents, type-safe structured outputs, typed dependency injection, OTel-native tracing, first-party durable execution — with an explicit API-stability promise the rest of the pre-1.0 field cannot match. The lowest-friction pick for teams that already think in Pydantic models.