OpenAI opened DevDay on Oct. 6, 2025 with a product aimed at a complaint its customers had made for a year2. Agent work sat scattered across too many separate tools. AgentKit put a visual builder, an embedded chat window, a connector registry and fine-tuning in one package. The adoption numbers arrived fast. Ramp cut its iteration cycles 70%, Canva saved two weeks of development time, and Carlyle lifted accuracy 30% while halving build time.
Twenty-three days earlier, Anthropic had renamed its own toolkit the Claude Agent SDK3. Four major runtimes shipped inside a year245. Read the pattern. Healthy competition settles on a standard. This looks more like every platform racing to become the layer a developer gets stuck in.
OpenAI's three-layer bet
Three levels make up OpenAI's approach, split on purpose. The Responses API sits lowest1. It is a bare interface for teams that want to “own the loop, tool dispatch, and state handling” themselves, fitted to short jobs that need little setup. Above it sits the Agents SDK. It wraps that API in a light Python framework built around handoffs, so one agent can pass a subtask to another. Guardrails check inputs and outputs before a user sees them. AgentKit sits on top, for teams that want a canvas rather than code2. Agent Builder offers drag-and-drop building with versions kept, and ChatKit handles the chat window a product team would otherwise write itself. Three levels, one company. The bet is that developers who start at the visual layer eventually need the code underneath, and owning both keeps them in the stack either way.
Anthropic extends what Claude Code proved
Where OpenAI built a new product, Anthropic extended one that worked. The company said so directly. "In other words, the agent harness that powers Claude Code (the Claude Code SDK) can power many other types of agents, too. To reflect this broader vision, we're renaming the Claude Code SDK to the Claude Agent SDK."3 It defined the result plainly too. "The Claude Agent SDK is a collection of tools that helps developers build powerful agents on top of Claude Code." Read the rename as evidence rather than marketing. A coding tool quietly became the engine behind deep research, video and note-taking. The name changed to catch up with the use.
Thariq Shihipar led the writing on that post, crediting seven other contributors3. Anthropic treats agent tooling as team engineering rather than one owner's project. Mike Krieger, the chief product officer, has spoken for the company's agent strategy in industry forums, including the Agentic AI Foundation launch covered elsewhere in this edition. The core claim carries weight. Claude Code proved the harness at developer scale before Anthropic sold it as a general framework, which handed the company a tested base rivals had to build or borrow.
Google and Microsoft build for their own clouds first
One lab moved earliest. Google unveiling its Agent Development Kit on April 9, 2025 at Google Cloud NEXT, authored by machine learning lead Erwin Huizenga and software engineer Bo Yang4. The post stated the goal in one line. "ADK is a new open-source framework from Google designed to simplify the full stack end-to-end development of agents and multi-agent systems." Its strongest credential predates the public release. The same framework already ran Agentspace and the Customer Engagement Suite inside Google before the code went open. Few competing frameworks can claim that at launch. LiteLLM support keeps ADK model-agnostic on paper. Its deepest tuning targets Gemini and Vertex AI, which ties the open framework's best speed to Google's own cloud.
Microsoft merged rather than built. It folded AutoGen and Semantic Kernel into one Agent Framework, with migration guides steering developers off both5. The project had 12.6 thousand GitHub stars and 2.1 thousand forks by this research's count. Its public docs leave product leadership unnamed. Chris DiBona, of Microsoft's Office of the CTO, has spoken for the company's open-agent strategy in nearby settings, including the Linux Foundation's Agentic AI Foundation.
The open-source layer holds its ground
LangChain built its position before any lab shipped a rival SDK, and the numbers show it. Harrison Chase and Ankush Gola's framework reaches 35% of Fortune 500 companies6. It has passed 1 billion open-source downloads, and its LangSmith tool takes in more than 1 billion events a day. That scale predates the current rush by years, built in the years before the major labs shipped anything of their own.
CrewAI, credited on its own blog to João Moura, claims 2 billion agentic workflows run through its platform and reports 65% Fortune 500 usage on its homepage7. It competes against the labs' own tools on the strength of a framework that stays independent of any one model provider. Mastra, a newer entrant built for TypeScript developers, has drawn 27.7 thousand GitHub stars. Fresh entrants keep arriving, even as the labs pour resources into competing runtimes.
LangChain says it works with 35% of the Fortune 500, and CrewAI claims 65%, each by its own count
Share of Fortune 500 companies each independent agent framework reports as users, by its own count, percent
Each bar is one framework's own claim about its share of the Fortune 500; the two claims come from separate counts, so each bar stands alone.
Sources: LangChain [6]; CrewAI [7]. Chart by Kelly Dennis.
The numbers behind this chart
| Item | Value | Note |
|---|---|---|
| LangChain, its own count | 35% | From LangChain's own about page, which also counts more than 1 billion open-source downloads |
| CrewAI, its own count | 65% | From CrewAI's homepage, which also claims 2 billion agentic workflows |
What the fragmentation implies for hiring and lock-in
Every runtime here builds toward the same neutral protocols this edition covers separately. Model Context Protocol handles tool access, and Agent2Agent handles agent-to-agent work1. The Linux Foundation governs both, rather than any one company. Neutral protocols settle the format two agents speak in. The harder question stays open.
A team fluent in LangGraph, or trained on the Claude Agent SDK's handoff patterns, pays real retraining cost to switch. That holds even when agents from each framework trade MCP calls cleanly. Hiring managers now screen for runtime-specific experience, the way earlier cycles screened for a cloud certification. That specificity is the barrier each vendor is building. Inference cost adds to it. Switching runtimes mid-deployment risks redoing cost work tuned to one framework, and few engineering leaders volunteer to pay that twice in a year.
Harrison Chase, the LangChain cofounder and CEO, sized up that competition for Fortune in October 2025. "There's a ton of players. I like to say we have 500 competitors and zero competitors at the same time."8 Chase expects most enterprises to run several agent platforms at once, with LangChain working underneath the ones OpenAI, Anthropic, Google, and Microsoft each built for themselves.
Jerry Liu, the LlamaIndex cofounder and CEO, reads the same crowd of rivals as proof the opposite is true. He told VentureBeat in May 2026 why demand for his own category keeps shrinking. "As a result, there's less of a need for frameworks to actually help users compose these deterministic workflows in a light and shallow manner," Liu said.9 Liu runs a framework that competes directly with LangChain, and he still named the layer both companies built as the one losing ground to models that plan and correct their own steps.
By the numbers
- Oct. 6, 2025: the day OpenAI launched AgentKit at DevDay, with Agent Builder, ChatKit and a connector registry2.
- Ramp reported a 70% cut in iteration cycles after adopting AgentKit2.
- Sept. 29, 2025: the date Anthropic renamed the Claude Code SDK to the Claude Agent SDK3.
- April 9, 2025: the date Google unveiled its Agent Development Kit at Cloud NEXT4.
- GitHub stars on Microsoft's Agent Framework stood at 12.6 thousand, after the AutoGen and Semantic Kernel merge5.
- LangChain reports working with 35% of Fortune 500 companies6.
- More than 1 billion open-source downloads, with LangSmith taking in over 1 billion events daily6.
- Two billion agentic workflows run through CrewAI's platform, by the company's own count7.
What to watch
Developer surveys that measure runtime share directly, rather than downloads or stars, would settle which framework wins production rather than tests. Compatibility layers, letting a team on one SDK call agents built on a rival's, would signal the industry choosing real interoperability over the lock-in each vendor now enjoys. Hiring data that tags runtime skill, the way cloud certifications once tracked AWS against Azure against GCP, would confirm whether agent work has become its own labor market.
Sources
- "OpenAI Agents SDK documentation," OpenAI, 2026, https://openai.github.io/openai-agents-python/.
- "Introducing AgentKit," OpenAI, Oct. 6, 2025, https://openai.com/index/introducing-agentkit/.
- Thariq Shihipar and others, "Building agents with the Claude Agent SDK," Anthropic, Sept. 29, 2025, https://claude.com/blog/building-agents-with-the-claude-agent-sdk.
- Erwin Huizenga and Bo Yang, "Agent Development Kit: Easy to build multi-agent applications," Google Developers Blog, April 9, 2025, https://developers.googleblog.com/en/agent-development-kit-easy-to-build-multi-agent-applications/.
- "microsoft/agent-framework," GitHub, Microsoft, 2026, https://github.com/microsoft/agent-framework.
- "About LangChain," LangChain, 2026, https://www.langchain.com/about.
- João Moura, CrewAI blog and homepage, CrewAI, 2026, https://www.crewai.com/blog.
- Sharon Goldman, "Exclusive: Early AI darling LangChain is now a unicorn with a fresh $125 million in funding," Fortune, Oct. 20, 2025, https://fortune.com/2025/10/20/exclusive-early-ai-darling-langchain-is-now-a-unicorn-with-a-fresh-125-million-in-funding/.
- Taryn Plumb, "The AI scaffolding layer is collapsing. LlamaIndex's CEO explains what survives.", VentureBeat, May 1, 2026, https://venturebeat.com/infrastructure/the-ai-scaffolding-layer-is-collapsing-llamaindexs-ceo-explains-what-survives.