The labs now say: use our harness, not a framework

Claim

The framework market split in two. Lab harnesses (Claude Agent SDK and Managed Agents, OpenAI Agents API, Google ADK and Managed Agents) own the loop. Independent frameworks (LangGraph, Pydantic AI, Vercel AI SDK, Mastra, Microsoft Agent Framework) converged on the same primitives: durable loop, tool-level approvals, middleware or hooks, OpenTelemetry. Role-play multi-agent frameworks (CrewAI, AutoGen) are being absorbed or sidelined; smolagents has had no release since May 2026.

Evidence

• Anthropic: Messages API (own loop), Agent SDK (runs the Claude Code binary with tools, hooks, subagents, sessions, skills), Managed Agents (hosted harness, beta header managed-agents-2026-04-01, primitives Agent, Environment, Session, Events; cron deployments; steer and interrupt; not ZDR or HIPAA eligible). Pricing of $0.08 per session-hour is secondary. UNVERIFIED.
• OpenAI: Agent Builder, the Evals platform and reusable prompt objects deprecated 2026-06-03, shut down 2026-11-30, with Promptfoo recommended for evals; Assistants API shut 2026-08-26; Agents API public beta 2026-09-10 (managed Codex harness, durable sessions, hosted or partner sandboxes, sandbox compute billed during active turns); GPT-6 Astra requires the Responses API for tool calling. Agents SDK is still 0.x (v0.24.0, 2026-10-15) and may break on minor releases.
• Google: ADK 2.0 moved from a hierarchical executor to a graph runtime (Python GA 2026-05-19, Go 06-30, TypeScript 08-21; 2.12.0 on 2026-10-13 with a task API and tool confirmation flow). Gemini CLI gained a policy engine, default plan mode, subagents and worktrees, and was replaced by Antigravity CLI for unpaid users on 2026-06-18 per its own changelog.
• Microsoft Agent Framework 1.0 (2026-04-03) merged Semantic Kernel and AutoGen: connectors, middleware, graph workflows with checkpointing, orchestration patterns, MCP; A2A "coming soon".
• Vercel AI SDK 7 (2026-06-25): durable WorkflowAgent, SandboxSession, OTel GenAI conventions, signed approvals, and HarnessAgent wrapping Claude Code, Codex and Pi. The TypeScript ecosystem now wraps lab harnesses rather than competing.
• LangGraph and LangChain 1.0 (2025-10-22): no breaking changes until 2.0; create_agent with middleware for approval, summarisation and PII. Pydantic AI 2.x: durable execution co-maintained with Temporal, DBOS, Prefect, Restate and AWS Lambda.

Open questions

• OpenAI's "open-source harness" and "no added fees" claims for the Agents API (announcement unreachable). UNVERIFIED.

Sources

• https://code.claude.com/docs/en/agent-sdk/overview — docs, Wed Oct 21
• https://platform.claude.com/docs/en/managed-agents/overview — docs, Wed Oct 21
• https://developers.openai.com/api/docs/changelog — docs, Mon Oct 19
• https://developers.openai.com/api/docs/deprecations.md — docs, Wed Oct 21
• https://openai.github.io/openai-agents-python/release/ — docs, Wed Oct 14
• https://adk.dev/2.0/ — docs, Wed Oct 21
• https://www.geminicli.com/docs/changelogs — docs, Mon Oct 12
• https://devblogs.microsoft.com/agent-framework/microsoft-agent-framework-version-1-0/ — blog, Thu Apr 02
• https://vercel.com/blog/ai-sdk-7 — blog, Wed Jun 24
• https://www.langchain.com/blog/langchain-langgraph-1dot0 — blog, Tue Oct 21
• https://pydantic.dev/docs/ai/capabilities/durable_execution/overview/ — docs, Wed Oct 21

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.