Open Source AI Agents: Frameworks Mature as Marketplaces Multiply

The short version

This fall, open-source AI agent software settled into a steady release schedule while the commercial business shifted to a different layer. Through late September and early October 2026, almost every major agent framework shipped an update within days of the others. Meanwhile security vendors, managed service providers, a crypto network and OpenAI itself opened marketplaces for buying, governing or paying agents. Taken together, the coverage points to one conclusion. The frameworks are becoming free, interchangeable infrastructure, and the money is moving to whoever controls distribution, governance and billing.

A crowded, fast-moving release calendar

Release logs make the pace clear. One comparison put LangGraph at version 1.2.13 on October 5. The same snapshot had CrewAI at 1.15.23 on September 28, OpenAI's Agents SDK at 0.23.1 on October 2, Google's ADK at 2.11.0 on October 2, and Pydantic AI at 2.54.0 on October 3.3 A week later another tracker had LangGraph at 1.2.14, CrewAI at 1.15.27 and Pydantic AI at 2.55.0.4 Google says ADK aims to ship roughly every two weeks.1

Several of these projects passed major-version milestones within the past year. LangChain 1.0 and LangGraph 1.0 shipped together on October 22, 2025, and the release made create_agent the center of the agent loop.5 Google's ADK 2.0 for Python arrived on May 19, 2026 with a graph-based workflow runtime. Pydantic AI reached version 2 on June 23, 2026 and can run agents durably on eight execution engines, including Temporal and DBOS.3

OpenAI's SDK started as the production successor to the experimental Swarm library in March 2025. In April 2026 it added native sandboxed execution and a "model-native harness" derived from the infrastructure behind Codex.10 One reference source notes that the SDK is built mainly around OpenAI's own models, and that support for other providers comes through a beta-level LiteLLM adapter.10 That detail matters for buyers who care about portability.

Consolidation is the real story in frameworks

The headline numbers matter less than the projects that disappeared or merged. Microsoft Agent Framework reached 1.0 on April 3, 2026 after a February release candidate. It combines AutoGen's orchestration patterns with Semantic Kernel's enterprise foundations and comes with a long-term support commitment.89 AutoGen itself is now in maintenance mode, with no new features and community management.9 Moving to the new framework requires real migration work: AutoGen's patterns were re-implemented rather than copied, so existing code has to be rewritten.5

The AutoGen lineage now runs in three directions. Microsoft's framework is the official successor, AutoGen 0.7.x is a maintenance line, and AG2 is a community fork that continues independently.12 Sources disagree on even basic facts here. One ranking dates AutoGen's latest release, 0.7.5, to September 30, 2025.1 Another lists the same version on September 30, 2026.12 The one-year gap changes how stale the project looks, and teams still running it should check the date themselves.

Consolidation has also reached tools developers rely on daily. Cursor acquired Continue, an early open-source coding assistant, around June 16, 2026. Its repository is now archived and read-only, and the last release shipped on June 19.5 Flowise's repository is also reported as archived and no longer receiving upstream fixes.12 Not every guide has caught up: a self-hosting roundup updated in October still recommends Flowise as the best no-code builder.11 Readers should check repository status, not just rankings.

The star-count wars, and why to discount them

GitHub stars still dominate coverage, and the numbers vary from source to source. OpenClaw, a self-hosted personal assistant that connects to WhatsApp, Slack, Telegram and other messaging apps, is described as the breakout project of 2026. One account says it went from about 9,000 stars to more than 382,000 by mid-year.1 Another counted 391,400 stars in early October.7 A third gives a rounder "300k+".11

The most striking claim concerns DeepSeek. One October analysis reports that the Chinese lab open-sourced "DeepSeek Harness" on August 13, 2026. It is described as an MIT-licensed TypeScript runtime where models, tools, skills and the UI all load as plugins, with roughly 248,000 stars, well ahead of LangChain's roughly 148,000.4 Other reporting confirms LangChain's count at about 147,500.7 However, the Harness figures appear in only one of the analyses reviewed here, so they should be treated as unconfirmed until other reporting backs them up.

The more useful point is the distinction that analysis draws between a framework and a "harness".4 LangChain now markets Deep Agents as its own harness, with planning, sub-agents, a filesystem and a sandboxed shell.9 OpenAI's SDK update used the same word.10 Competition is moving away from how you write an agent loop and toward how much finished runtime comes with it.

Several comparisons warn that stars measure attention, not quality or production readiness.12 That warning is easy to agree with. Some of these rankings are also published by vendors that place their own product first, including Sim, Gumloop and Munder Difflin.121320

"Open source" increasingly comes with an asterisk

Licensing is getting more complicated. CrewAI's core is MIT-licensed, but its visual editor and governance features sit in the hosted AMP layer, which has no stated open-source license.14 Mastra and Sim pair Apache 2.0 cores with separately licensed enterprise directories.312 n8n uses the Sustainable Use License, which is not OSI-approved. Dify adds multi-tenant conditions on top of Apache 2.0.12 AutoGPT's platform code is under Polyform Shield, while its classic code stays MIT.7

The pattern is consistent: the framework is free, and you pay for tracing, deployment and control. One guide puts it plainly. LangChain, LangGraph, Microsoft Agent Framework and CrewAI's core cost nothing, and the bills come from models, compute and hosted layers.9 Telemetry is another thing to check. One comparison notes that the OpenAI Agents SDK sends traces to OpenAI by default and that CrewAI turns on anonymous telemetry by default.3

Interoperability standards are increasingly what hold this together. Google created Agent2Agent (A2A), which is now an Apache-licensed Linux Foundation project. Its steering committee includes AWS, Cisco, IBM Research, Microsoft, Salesforce, SAP and ServiceNow, and version 1.0.1 shipped on May 28, 2026.7 Microsoft's framework supports MCP, A2A and AG-UI.8 Goose has moved from Block to the Agentic AI Foundation at the Linux Foundation.20

Marketplaces: where the money is moving

While frameworks became more alike, marketplaces multiplied. The common thread is that each one changes how agents are bought and billed, not how they are built.

Netskope launched its AgentSkope Marketplace on October 6. It replaces buying agents one at a time with a single catalog, a shared pool of capacity credits, a credit calculator and a usage dashboard.21 The company cited a survey in which 42% of security and network operations leaders said alerts or tickets went uninvestigated.22 Trade coverage described the change as mainly commercial and administrative, not technical.28 That matches the reading that procurement friction, more than capability, is holding back enterprise agent adoption.

Kiteworks went further on governance. Its marketplace offers more than 60 agents that install into Claude and route through the company's Secure MCP Server. Credentials stay out of the model's context, and every action is tied to an authenticated user.27 Kiteworks also publishes each agent's logic as readable definitions that customers can review.27 In effect, it uses openness as an audit feature rather than a licensing model.

OpenAI's new B2B Marketplace may be the most significant change in distribution. Sierra joined as a launch partner, so eligible enterprises can put existing OpenAI spending commitments toward Sierra agents, which are priced on outcomes such as a closed sale or a retained customer.29 One analyst called it a deliberate distribution strategy that puts Sierra inside procurement workflows enterprises already use.29 That reading looks right. When model vendors become the payment channel, independent agent companies have to compete for placement inside them.

At the edges, Celo Core Co. launched "buy" on October 13. It lists more than 700 compute, data and API services that agents can pay for themselves in stablecoins, with no human creating an account or API key.23 Access to some services, including Google Cloud's largest machines, requires a zero-knowledge proof that a verified adult is behind the agent.26 In the UK, managed service provider Air IT started an agent marketplace for SMEs that deploys pre-built agents into customers' Microsoft environments.

What it adds up to

The evidence supports a clear conclusion. Open-source agent frameworks are now well-documented, frequently released and mostly interchangeable, and the large vendors either fund them or fold them into their own products. Consolidation by Microsoft, Cursor and others is cutting the number of independent options, while open-core licensing pushes the parts that make money into proprietary layers.

The new marketplaces show where value is collecting: governance, procurement and payment. For developers, the practical advice is to choose frameworks by maintenance status and standards support, not star counts. For enterprises, the next fight is over who controls the catalog and the bill, not which framework wins.