AI Agent Frameworks Converge on MCP as 2026 SDK Releases Pile Up

Every major lab now ships an agent SDK

In 2026 the agent-framework market turned into a contest between platform vendors. LangChain no longer has the field mostly to itself. OpenAI, Google, Microsoft and Anthropic have all released official agent SDKs, and several comparison guides call this the main change of the year.28 A widely circulated October tracker covers eight SDKs: Claude Agent SDK, OpenAI Agents SDK, Google ADK, LangGraph, CrewAI, Smolagents, Pydantic AI and Microsoft Agent Framework 1.0. It also includes a primitive-by-primitive reference for Anthropic's toolkit, covering the query() entry point, lifecycle hooks, subagents, streaming, isolation and MCP handling.21

The release pace is the first thing that stands out. In the first eleven days of October 2026 alone, new versions came out for the Claude Agent SDK (Python 0.2.165, TypeScript 0.3.296), OpenAI Agents SDK (Python 0.23.1), Google ADK (2.11.0), LangGraph (1.2.14), CrewAI (1.15.27), Pydantic AI (2.55.0), Microsoft Agent Framework (1.21.0) and Mastra (1.76.0).1 The outlier is AutoGen. Its last listed release dates to September 2025.1

My view is that framework choice now matters less than it did a year ago. The thing that ties these products together is a shared protocol layer, the Model Context Protocol (MCP), and that layer has just been rebuilt.

The consolidation year: Microsoft, Google and OpenAI hit milestones

Coverage agrees that Microsoft made the biggest move. It merged AutoGen and Semantic Kernel into Microsoft Agent Framework, released a release candidate on February 19, 2026, and declared 1.0 general availability in early April with native MCP and A2A support for Python and .NET.2123 One analysis calls it the most consequential release of the year because it folded Microsoft's two most-used agent SDKs into one.23 Sources disagree slightly on the date: most give April 3, while one tracker that checks GitHub repositories says April 2.2126 AutoGen's README now says the project will get no new features and sends users to the new framework.26 A Go SDK is in public preview.

Google's path ran through more languages and a new architecture. ADK Java 1.0 and Go 1.0 shipped early in the year.21 ADK 2.0 for Python then arrived on May 19, 2026, with a graph-based workflow runtime that handles routing, fan-out and fan-in, loops and retries, plus a Task API for delegating work between agents.26 One reviewer notes that this is the same move LangGraph made years earlier, now combined with native Gemini integration.25 Most listings now count five supported languages, including Kotlin.26

OpenAI overhauled its Agents SDK in April 2026, adding native sandbox execution, sub-agent patterns, filesystem tools and first-class MCP support.22 One tracker says the sandbox harness works with seven providers, including Cloudflare, E2B, Modal and Vercel.23 Even so, the SDK's version number is still below 1.0, and one October review raises that as a caveat next to frameworks that have formally reached 1.0.26

Anthropic's SDK: deepest MCP, most disputed details

The Claude Agent SDK has a clear identity but draws more disagreement than any other framework here. Anthropic renamed it from the Claude Code SDK, signalling ambitions beyond coding.26 Reviewers broadly agree that it treats the agent as a stateful runtime with deep operating-system access, shipping built-in tools for reading, writing and editing files, running Bash, searching code and fetching web content.3 It also has the deepest MCP integration of the group, which makes sense because Anthropic created the protocol.16 The usual recommendation is to use it for coding agents and other computer-use work.19

After that, the reporting splits. On licensing, one comparison calls the SDK proprietary, while a hands-on test lists it as MIT-licensed.37 On state, one analysis says sessions resume automatically, while another says state lasts only for the session and a crash means starting over.46 On maturity, a guide from March 2026 rated it alpha.27 By the second half of the year, other reviewers were giving it top production scores, partly because Anthropic's own Claude Code runs on it.23

Cost is the most concrete criticism. In one test using the same model and question ten times per framework, the Claude Agent SDK averaged about 35,000 input tokens and an 8.5-second median latency. The OpenAI Agents SDK averaged roughly 1,350 tokens and 2.5 seconds. Both answered correctly every time.7 That overhead is the price of shipping a full harness. Teams building simple agents should take it seriously, while teams that need file and shell access may find it worth paying. Commercial terms are also moving: a billing change for the SDK was announced and paused on the same day in June.21

The model lock-in point is not in dispute. The SDK works only with Claude models, while OpenAI's SDK reaches more than 100 models through LiteLLM.324

MCP becomes stateless, and that matters more than any single SDK

The most important development under these frameworks happened in the protocol. The MCP 2026-07-28 specification shipped on schedule, and its maintainers describe it as the largest revision since launch.1114 It removes the initialize handshake and the Mcp-Session-Id header. Each request now carries its own protocol version, client identity and capabilities, so it can be handled by any server instance behind an ordinary round-robin load balancer.14 Method and tool names move into HTTP headers so gateways can route and authorize traffic without parsing request bodies. Tool lists come with cache hints, and Dynamic Client Registration is formally deprecated in favour of client metadata documents.14 Tasks and MCP Apps, which let servers render interactive interfaces in sandboxed iframes, are now formal extensions.11

Infrastructure vendors responded quickly. Cloudflare says MCP servers can now run in a single Worker with no stateful infrastructure, and that the TypeScript, Python, Go and C# SDKs shipped alongside the spec.12 Google called the release a turning point that takes MCP from a local integration layer to enterprise infrastructure.18 Manufact, which hosts thousands of MCP servers, said the new SDK cut its package size by about 83 percent and made it 25 percent faster.14 The change also breaks things: Roots, Sampling and Logging are deprecated, and a 12-month window for legacy versions is now running.1620 One German-language analysis cites The New Stack's point that the protocol has stopped managing state, but the state has not disappeared. Servers that need application state now have to handle it explicitly.19

The maintainers' August roadmap sets out the next priorities: unifying on HTTP-native transport, agentic messaging primitives, agent identity, and a better SDK developer experience.15

Scale, security and the emerging MCP marketplace

The ecosystem figures help explain why these changes matter, although they come from second-hand aggregations and should be read with care. One roundup counts about 15,930 public MCP servers across four major registries, 9,652 of them in the official registry, plus more than 97 million monthly SDK downloads.20 The same roundup cites independent scans that found exploitable flaws in somewhere between 30 and 82 percent of public servers, depending on the scan. It also reports that the NSA and CISA published joint security design guidance on June 2, 2026.20

On the commercial side, MCP servers are becoming a standard feature of SaaS products. Front launched an MCP server on October 6 so external AI assistants could connect to customer-operations accounts.13 In the same week, NinjaOne released an MCP connector for IT operations, and Scytale offered scoped, read-only MCP access to compliance data.17 The pattern is that vendors now see an MCP endpoint as a way to get their data in front of Claude, ChatGPT and similar clients, much as a public API once was.

What it means

The coverage leads to a simple conclusion. The choice between agent SDKs mostly comes down to existing stack loyalty: .NET and Azure teams go to Microsoft, multi-language enterprises to Google ADK, stateful workflows to LangGraph, coding and computer-use agents to Anthropic, and quick, model-flexible prototypes to OpenAI.16 Every one of them now supports MCP in some form, which one guide describes as a shared integration protocol underneath the orchestration layer.10

The open questions sit in that protocol layer, not in the frameworks. They include how quickly legacy MCP deployments migrate, whether the security gap narrows before the 2027 deprecation deadline, and whether agent-to-agent standards such as A2A, currently native mainly in Google's and Microsoft's frameworks, become as widespread as MCP.123 The frameworks will keep shipping versions every week. Where the market ends up depends more on the protocols underneath them.