Top GitHub Repositories for AI Agentic Frameworks and Document Extraction — Mid-2026 Check-In (Part 2)

Back in February 2026, we mapped the open-source landscape for AI agentic frameworks, enterprise agentic platforms, and document extraction tools. Six months later, it’s time for a check-in: what held steady, what consolidated, and what’s newly rising. This is Part 2 of a three-part series — we’ll do a final comparison at the end of 2026 to see how the full year played out.

Why Revisit So Soon

Six months is a short window, but 2026 has been anything but quiet for this space. The biggest story isn’t a new repo climbing the charts — it’s consolidation. Microsoft merged two of the frameworks from our original list into one. Protocols that were emerging in February are now table stakes. And the extraction layer keeps compounding stars faster than the agent layer itself.

Part 1 — What Changed in the Agentic Core

The Big Consolidation: AutoGen + Semantic Kernel → Microsoft Agent Framework

The single biggest structural change since February: microsoft/autogen and microsoft/semantic-kernel — both featured separately in our original post — are now in maintenance mode. Microsoft merged them into a single unified SDK, microsoft/agent-framework, which hit GA (1.0) on April 3, 2026. It combines Semantic Kernel’s plugin/connector model with AutoGen’s multi-agent orchestration into one graph-based workflow engine, with native MCP and A2A support across .NET and Python. If you picked either framework in February, official migration guides now point you to the new repo — this is the clearest signal yet that the “pick a framework and commit” calculus has shifted toward frameworks that treat protocol interoperability as a first-class feature, not an integration.

Star Growth: Steady, Not Explosive

LangChain remains the biggest single repo in the space, but growth has slowed to a steady climb rather than the hockey-stick trajectory of 2024–2025 — a sign the core framework layer is maturing rather than churning.

Protocol Standardization: MCP Won, A2A Joined It

In February, we called out “MCP everywhere” as a trend to watch. By mid-2026, that trend has fully materialized:

• MCP (Model Context Protocol) is now supported natively — not via community plugin — in every framework from our Tier 1–3 lists: LangGraph, Google ADK, Mastra, OpenAI Agents SDK, and the new Microsoft Agent Framework.
• A2A (Agent-to-Agent) protocol has emerged as MCP’s companion, standardizing how independent agents (potentially built on different frameworks) discover, negotiate, and hand off tasks to each other. Frameworks now routinely advertise “MCP + A2A” support as a checkbox feature.
• MCP’s governance has moved toward the Linux Foundation, with 200+ community server implementations — a strong signal it has crossed from “Anthropic’s idea” to genuine industry standard.

What this means for framework choice: the “how do I connect tools to my agent” question that used to differentiate frameworks is now mostly solved and portable. Differentiation is shifting to observability, guardrails, and multi-agent state management — exactly the areas Google ADK, Mastra, and PydanticAI have been investing in.

Rising Names to Watch

A few projects that weren’t prominent (or didn’t exist) in our February list are showing up in mid-2026 roundups:

• Hermes Agent — a cross-provider agent harness with deep MCP integration, oriented around team/messaging workflows. (We covered it head-to-head with OpenClaw in a separate post.)
• Browser Use and Firecrawl — narrower, browser/web-context-focused agent tools gaining traction as “the retrieval layer” for agents that need live web data rather than static documents.

Neither has cracked the 20k+ star tier yet, but both are worth tracking for the year-end comparison.

Part 2 — Enterprise Platforms: Steady Leaders

The enterprise tier from February held up well — no major reshuffling:

• langgenius/dify and n8n-io/n8n remain the two dominant “full platform” plays, both comfortably past the 50k mark, with n8n’s workflow-automation angle continuing to pull in teams that want agents wired directly into existing business processes.
• BerriAI/litellm keeps benefiting from the multi-provider reality enterprises live in — as more frameworks add first-class multi-model support, a unified gateway layer is still valuable for cost tracking and governance rather than routing logic itself.
• microsoft/semantic-kernel, as noted above, is now the “old” enterprise SDK — still widely deployed, but new projects are steered toward Microsoft Agent Framework.

No new enterprise-tier entrant has displaced this group yet, reinforcing that enterprise adoption cycles move slower than the raw framework layer — teams that picked a platform in 2025 are mostly still on it.

Part 3 — Document Extraction: Still the Fastest-Growing Layer

If Part 1 was about consolidation, Part 3 is about acceleration. The extraction layer is growing faster than the agent layer, confirming the trend we called out in February (“extraction quality is the new bottleneck”):

MarkItDown’s jump is the single largest star-growth story across our entire original list — it has effectively become the default “throw any file at an LLM pipeline” utility, now claiming ~82% F1 accuracy on standard extraction benchmarks and adoption in 2,700+ production projects.

What’s new under the hood:

• MinerU v3.4.0 shipped a major OCR upgrade (PP-OCRv6, +11% accuracy), hybrid speed/accuracy backends, and support for 10+ non-mainstream AI chip architectures — a sign extraction tooling is being pulled into on-prem and sovereign-AI deployments, not just cloud pipelines.
• Docling released a lighter “docling-slim” build alongside a new docling-parse v7 backend for better table and layout inference, plus expanded ASR (audio-to-document) support — extraction is quietly absorbing audio/video sources, not just paper documents.
• MarkItDown now processes ~100 pages in ~12 seconds with no GPU required, reinforcing why it remains the lowest-friction entry point for teams bolting extraction onto an existing agent stack.

The smaller Tier 2/3 tools from our original list — Unstructured, ContextGem, DocStrange — haven’t seen the same explosive growth, suggesting the market is consolidating around a few extraction leaders (MarkItDown, MinerU, Docling) the same way the agent layer consolidated around LangChain/LangGraph, CrewAI, and now Microsoft Agent Framework.

Updated Trends to Watch (for the End-of-Year Comparison)

1. Protocol consolidation is done; framework consolidation is next. MCP/A2A adoption is now universal. Watch instead for further framework mergers or sunsettings, following Microsoft’s lead with AutoGen + Semantic Kernel.
2. Extraction is compounding faster than orchestration. MarkItDown’s +51k stars in six months dwarfs any agent framework’s growth. By year-end, expect the extraction tier to rival or exceed the agent tier in raw popularity.
3. On-prem and sovereign AI is pulling extraction tooling forward. MinerU’s chip-architecture support signals extraction is now a target for regulated/on-prem deployments, not just SaaS RAG pipelines.
4. Audio and multimodal extraction is creeping in. Docling’s ASR support hints that “document extraction” is quietly becoming “any-media extraction.”
5. Enterprise platform choices are sticky. Dify, n8n, and LiteLLM haven’t been meaningfully challenged — expect the next disruption here, if any, to come from a new entrant rather than incremental share-shifting among 2025’s leaders.

What We’ll Check at Year-End (Part 3)

For the final 2026 comparison, we’ll specifically track:

• Whether Microsoft Agent Framework overtakes standalone AutoGen/Semantic Kernel usage in real deployments, not just stars.
• Whether MarkItDown’s growth curve holds, or whether Docling/MinerU close the gap on accuracy-sensitive use cases.
• Whether any framework not on our radar today (Hermes Agent, Browser Use, or something not yet released) breaks into the 20k+ tier.
• Whether A2A gets formalized under a governance body the way MCP has, or fragments into competing standards.

Conclusion

Six months in, the biggest lesson isn’t a new repo to bookmark — it’s that the ecosystem is consolidating around fewer, broader tools at every layer: one merged Microsoft framework instead of two, a small handful of extraction leaders pulling away from the pack, and protocols (MCP, A2A) that used to be differentiators now expected by default. Pick your stack with an eye toward which side of that consolidation a tool is on. We’ll revisit this list one more time in December to see how the year closed out.

Star counts reflect approximate values as of late June/early July 2026, gathered from public GitHub statistics and community trackers. Follow my blog for the final 2026 comparison in this series.