Skip to content

The AI Agent SDK Landscape — Popularity, Design Philosophy, and Downstream Projects (Aug 2026)

Research date: Aug 29, 2026. Star counts were pulled live from the GitHub API; Python download counts are the pypistats 30-day rolling window; npm counts are the npm registry's last-full-month totals (all fetched Aug 28–29, 2026). Ecosystem and community claims are sourced inline. This is a popularity, philosophy, and adoption snapshot — not a capability benchmark.

Companion post: LangGraph in 2026 — Deep Dive and Community Views.

Part 1 — Popularity

Short answer

OpenAI Agents SDK is the more popular of the two by most measures, and the gap widened a lot in 2026. It has more GitHub stars (29.0k vs 19.6k), ~3x the downloads of the main Python package (38.8M vs 12.6M per month), a TypeScript SDK that Pydantic AI lacks, and a much higher momentum score on the daily framework tracker (79.0 vs 54.1).

But "popular" depends on the metric, and the honest answer is more interesting:

  • Pydantic AI's full Python install surface is much larger than its main package suggests. Its core distribution, pydantic-ai-slim, pulled 132.6M downloads/month (the batteries-included pydantic-ai package is 12.6M and depends on slim), so on raw Python install volume the two are far closer — possibly Pydantic AI ahead.
  • Pydantic AI wins the supply-chain/trust ranking (HVTracker #2 vs OpenAI Agents SDK #4).
  • As recently as March 2026 the two were nearly tied on PyPI downloads (~14M vs ~13M). OpenAI pulled away during H1 2026.
  • Neither is the overall ecosystem leader. LangGraph has the most production/state-machine adoption and the highest agent-SDK download volume (69.3M/month); CrewAI has the most stars among "pure" agent frameworks (57.7k); Vercel AI SDK is the dominant TypeScript layer (90.7M npm downloads/month); BrowserUse is the hottest project by momentum.

Head-to-head: Pydantic AI vs OpenAI Agents SDK

Metric OpenAI Agents SDK Pydantic AI
GitHub stars (Aug 28, 2026) 29,038 (openai-agents-python, created Mar 2025) 19,554 (pydantic-ai, created Jun 2024)
TypeScript story openai-agents-js 3.7k stars; @openai/agents ~6.2M npm downloads/mo None — Python-only
PyPI downloads/mo (30d to Aug 28) openai-agents 38.8M pydantic-ai 12.6M + pydantic-ai-slim 132.6M
Momentum, Pulse Score (ai-agent-pulse, Aug 27) 79.0 (#3 of 19) 54.1 (#14 of 19)
Trust/supply-chain rank (HVTracker, May 30) #4 (90.2) #2 (90.6)
PyPI downloads/mo, March 2026 (quark-agents) ~14M ~13M
Latest releases (Aug 2026) v0.21.0 (Aug 15) v2.35.x (Aug 28)

How to read this. Stars and momentum favor OpenAI clearly: ~29k stars in ~18 months versus ~19.6k in ~26 months, with far higher commit activity in the pulse tracker. PyPI downloads are messier — pydantic-ai depends on pydantic-ai-slim, so the packages overlap; the clean single-package comparison favors OpenAI ~3:1, while the full Pydantic AI install surface is on the order of 130M+ monthly downloads. Both figures include CI/bot/mirror noise, so treat them as deployment-intensity proxies, not unique-developer counts.

Why OpenAI pulled away: distribution gravity (OpenAI docs, ChatGPT/GPT ecosystem, built-in tracing, sandbox agents), Python and TypeScript support, and path-of-least-resistance defaults for the huge OpenAI user base. Why Pydantic AI still wins fans: the Pydantic brand, excellent type safety and structured-output validation, native async/streaming, Logfire observability, dependency injection, and a minimal-dependency design — plus the higher verifiable trust signals (provenance, signed commits, OSSF Scorecard).

The other alternatives (Aug 2026)

Framework Languages Stars (Aug 28) PyPI/mo npm/mo Niche
LangGraph Python, TS 40.6k 69.3M (langgraph) 13.2M (@langchain/langgraph) State machines, checkpointing, enterprise production (Klarna, Uber, LinkedIn)
CrewAI Python 57.7k 30.5M — Role-based teams, fastest prototyping; ~2B agent executions/yr claimed; NVIDIA partnership
Claude Agent SDK Python, TS 8.0k (SDK); Claude Code repo 143.3k 32.6M (claude-agent-sdk) 40.5M (@anthropic-ai/claude-agent-sdk) Anthropic-native; deepest MCP integration; coding/OS agents; TS package bundles Claude Code
Google ADK Python, TS, Java, Go, Kotlin 21.3k 21.3M (google-adk) — Google-native, 5 languages, A2A-native; ADK 2.0 GA May 19, 2026
Microsoft Agent Framework Python, .NET 13.2k 7.5M (azure-ai-agents) — MAF 1.0 GA Apr 3, 2026; successor to AutoGen + Semantic Kernel; MCP + A2A native
Vercel AI SDK TypeScript 26.5k — 90.7M (ai) + 43.5M (@ai-sdk/openai) Dominant TS provider abstraction
Mastra TypeScript 27.5k — 2.6M TS-native agent framework; most active dev (1690 commits/4w)
smolagents Python 29.0k 0.6M — Hugging Face; lightweight code-executing agents
Agno Python 42.0k 2.2M — Formerly Phidata; tool-rich multi-agent
AutoGen / AG2 Python 60.7k small (autogen-agentchat) — Maintenance mode; no releases since Sep 2025
LlamaIndex Python 51.9k — — RAG/data-first agents
BrowserUse Python 111.6k — — Web agents; hottest momentum (+1612 stars/wk)
Semantic Kernel C#, Python, Java ~28.5k — — Enterprise .NET/Java; maintenance, successor is MAF

Honorable mentions: DSPy (prompt/loop optimization, 37.6k), Haystack (26.3k), AgentScope (Alibaba, ~29.9k), Letta (agent memory), Composio (tool integrations), OpenManus (56k), Strands Agents, Instructor, and ControlFlow (archived under Prefect). On the managed side, Vertex AI Agent Engine, AWS Bedrock Agents, and Azure AI Foundry are the vendor-hosted runtimes behind the SDKs above.

The trackers: is there a benchmark ranking AI SDKs?

Several exist, and they measure different things:

  1. ai-agent-pulse — the closest thing to a daily popularity leaderboard: a GitHub Action recomputes a Pulse Score (0–100) for 19 frameworks from star velocity, release freshness, commit activity, issue health, community size, and fork engagement. Aug 27, 2026: BrowserUse 89.4, CrewAI 81.0, OpenAI Agents SDK 79.0 (#3), Mastra 78.0, Google ADK 74.5, LangGraph 73.7, Claude Agent SDK 70.0, PydanticAI 54.1 (#14).
  2. AgentMarketCap — composite rankings of 500+ agents/frameworks plus analysis posts.
  3. HVTracker — ranks by trust/supply-chain signals (provenance, OSSF Scorecard, signed commits): LangGraph #1, PydanticAI #2, OpenAI Agents SDK #4 (May 2026).
  4. Presenc AI SDK Download Rankings — quarterly PyPI + npm download leaderboards for vendor SDKs (May 2026: LiteLLM ~418M, OpenAI ~380M combined, Google ~275M, Anthropic ~191M).
  5. quark-agents framework comparison — one-off March 2026 ranking by PyPI downloads plus source-line analysis.
  6. OSSInsight / Star History — raw star/commit/PR trend curves.

Do not confuse these with capability benchmarks (SWE-bench, AgentBench, GAIA), which measure task performance, not developer adoption. Popularity trackers are proxies: downloads include CI/mirror noise, stars include inactive users, and pulse scores use different weights per tracker.

Ecosystem snapshot — Aug 2026

Consolidation is the story of the year. Microsoft merged AutoGen and Semantic Kernel into Microsoft Agent Framework 1.0 (GA April 3, 2026), and AutoGen's README carries a maintenance-mode notice with no release since September 2025. LangChain and LangGraph settled into stable 1.x lines (LangChain's community-integrations repo was archived in June 2026), and Thoughtworks moved LangGraph from Adopt to Trial in April 2026.

Every major lab ships a first-party SDK now. OpenAI Agents SDK (v0.21.0, Aug 15, 2026) added sandbox agents, sessions, and built-in tracing; Claude Agent SDK (renamed from the Claude Code SDK in early 2026) centers on query(), hooks, and the deepest MCP integration, with a June billing change announced and paused the same day; Google ADK 2.0 went GA May 19, 2026 with graph-based workflows, five languages, and native A2A; Microsoft's MAF 1.0 rounds out the four.

The protocol layer is shrinking lock-in. MCP was donated to the Linux Foundation in December 2025 (200+ server implementations); A2A passed 150 organizations by April 2026, and ACP merged into A2A under the Linux Foundation. Tool access and cross-agent communication are increasingly protocol-level concerns.

The routing/abstraction layer is quietly the biggest install surface. LiteLLM hit ~707M monthly PyPI downloads in August 2026 — an order of magnitude above any individual agent SDK — and Vercel AI SDK dominates TypeScript (ai at ~91M npm/month plus provider packages). The practical 2026 stack is three-tier: vendor SDK + orchestration framework + gateway.

Part 2 — Design philosophy

The framing

Every framework is a bet on five questions: who controls the loop (developer or model), what the unit of abstraction is (agent, role, graph node, workflow, code), who owns state (framework, vendor, or nobody), how it relates to the model vendor (native, agnostic, abstraction layer), and what "good" means by default (type safety, speed, durability, observability, portability, simplicity). On those axes the field splits into three camps:

  • "The agent loop is a commodity" — OpenAI Agents SDK, Claude Agent SDK, Vercel AI SDK: tiny primitives; value is the model, tools, and observability.
  • "Orchestration is the product" — LangGraph, CrewAI, Mastra, Google ADK, Microsoft Agent Framework: coordination, state, and durability are the hard part, so the framework owns them.
  • "Correctness is the product" — Pydantic AI, DSPy, smolagents: typed outputs, validation, and optimization matter more than orchestration.

The bets, framework by framework

  • OpenAI Agents SDK — two driving design principles, verbatim from the docs: "enough features to be worth using, but few enough primitives to make it quick to learn" and "works great out of the box, but you can customize exactly what happens." Python-first: orchestrate with ordinary language features, not new abstractions. The loop is a solved problem; value is speed + tracing + (now) sandboxes and sessions.
  • Pydantic AI — "typed end to end"; outputs are validated Pydantic models, tool schemas come from type hints, the same agent runs behind web/CLI/voice/queue. Deliberately no built-in memory and no orchestration-graph default. The lineage is Pydantic/FastAPI: powerful tools that are still easy to use.
  • LangGraph — "control and durability, so you have the best chance of your agent doing what you intended" (its own first-principles blog). Agents are state machines: nodes, edges, checkpointed state, time-travel, human-in-the-loop interrupts. Explicit by construction; you pay design cost up front. (See the LangGraph companion post for 2026 evolution and the community verdict.)
  • CrewAI — agents are roles ("don't build one agent, build a crew"): role, goal, backstory as structural elements. Crews for autonomy, Flows for deterministic production control. Token-hungrier than code-routed orchestration.
  • Claude Agent SDK — "give the agent a computer": one primitive (query()), real OS tools out of the box, permission modes and hooks for governance, deepest MCP integration, deliberately Claude-only. Anthropic's meta-philosophy (Building Effective Agents): start simple, avoid abstraction layers, workflows vs agents.
  • Google ADK — "applies software development principles to AI agent creation": code-first, modular, model-agnostic, deployment-agnostic; ADK 2.0 is graph-based with five languages and native A2A. Agents as well-engineered software.
  • Microsoft Agent Framework — production-grade, governance-first: durability, restartability, observability, human-in-the-loop, declarative YAML workflows, native MCP + A2A. AutoGen's conversation-as-coordination survives as a pattern.
  • Vercel AI SDK — provider-agnostic TS toolkit: unified API, streaming primitives, UI hooks, gateway by default. Deliberately not an agent runtime; portability and frontend DX are the product.
  • Mastra — TS-first, convention over configuration with progressive complexity: Agents (emergent) + Workflows (explicit .then()/.branch()/.parallel()), memory, suspend/resume, OTel.
  • smolagents — agents that think in code: the model writes Python instead of JSON tool calls; the agent logic fits in ~1,000 lines. Simplicity and action expressiveness over machinery.
  • AutoGen — conversation as coordination: agents solve tasks by talking to each other. Retired, but the pattern shaped the category and lives on in MAF.
  • DSPy — "programming, not prompting": prompts are programs you compile and optimize, not strings you hand-tune.

The axes, in one table

Framework Control vs autonomy Unit of abstraction State ownership Model relationship Default good
OpenAI Agents SDK Managed loop, configurable Agent + handoffs Sessions (SDK/vendor) Agnostic-ish, OpenAI-optimized Speed + tracing
Pydantic AI Typed agent loop Typed Agent You manage Agnostic Type safety + DX
LangGraph Explicit control Graph nodes/edges Checkpoints (framework) Agnostic Durability + auditability
CrewAI Autonomy by default; Flows for control Crews / roles Limited Agnostic Multi-agent DX
Claude Agent SDK Autonomous with permission gates Agent + subagents + query() Sessions (vendor) Claude-only Capability + safe tool access
Google ADK Code-first workflows Agent tree / graph Sessions (framework) Gemini-optimized, agnostic Enterprise multi-language
Microsoft Agent Framework Graph + conversation patterns Workflow + agents Checkpoint storage Agnostic Governance + durability
Vercel AI SDK Transport, not orchestration generateText / providers Your app Agnostic (gateway) Portability + frontend DX
Mastra Agents + explicit Workflows Agent / Workflow Storage (framework) Agnostic (40+ providers) TS DX, prototype→production
smolagents Autonomous in code CodeAgent Conversation memory Agnostic Simplicity + code as action
DSPy Programmatic (compile-time) Module / Signature Your app Agnostic Systematic optimization

Synthesis

Features are converging — OpenAI added sessions and sandboxes, ADK 2.0 became graph-based like LangGraph, CrewAI recommends Flows for production, Mastra ships workflows, Pydantic AI added a graph API. The first 100 lines differ a lot; the 10,000-line experience differs much less. Protocols (MCP, A2A) are de-risking interop, so the remaining lock-in is architectural — your state schema, graph shape, and vendor sessions — not legal.

Part 3 — Downstream projects: who actually builds on these SDKs

The three builder types

  1. The labs themselves ship reference apps and samples: Vercel's ai-chatbot, Google's adk-samples, Microsoft's agent-framework-samples, Anthropic's claude-cookbooks, LangChain's open-canvas/deepagents, Hugging Face's agents-course.
  2. Big companies build branded tools on the mainstream SDK: Coinbase's AgentKit on the OpenAI Agents SDK (cited in OpenAI's launch post); Klarna, Uber, LinkedIn, Replit, Elastic on LangGraph (mostly private).
  3. Community and solo devs ship viral utilities: TradingAgents (Tauric Research), GPT Researcher (Assaf Elovic), claude-mem, Shannon (KeygraphHQ), oh-my-claudecode, nanoclaw, pydantic-deepagents.

The downstream map (verified builds)

Framework Notable public downstream Builder Stars (Aug 2026)
OpenAI Agents SDK AgentKit Coinbase (CDP team) 1.3k
PageIndex — first-class SDK integration VectifyAI 35.4k
agents — coding/computer-use/deep-research agents Lutz Roeder community
durable-agentic-harness Temporal demo
Pydantic AI pydantic-deepagents — self-hosted "open-source Claude Code" VStorm Co community
LangGraph TradingAgents — multi-agent trading Tauric Research 101.6k
GPT Researcher Assaf Elovic 29.2k
Deep Agents LangChain 28.7k
Open Canvas LangChain 5.5k
CrewAI crewai-examples CrewAI Inc 6.1k
Claude Agent SDK / Claude Code claude-mem thedotmack 92.5k
Shannon — AI pentester KeygraphHQ 47.3k
oh-my-claudecode Yeachan-Heo 38.8k
nanoclaw nanocoai 30.6k
Google ADK adk-samples — 6 languages Google 10.2k
Microsoft Agent Framework agent-framework-samples Microsoft 366
Vercel AI SDK ai-chatbot Vercel ~16k+ (2025 snapshot)
chatbot-ui McKay Wrigley 33.3k (stale)
Mastra examples + personal-assistant-example Mastra team official
smolagents agents-course Hugging Face 31.7k

Notes and caveats

  • Every entry was verified against the repo's README and, where possible, its dependency manifest — a README mention is not proof of use.
  • AgentKit and PageIndex are integrations as much as pure downstreams (AgentKit also ships LangChain bindings).
  • Most of the Claude-ecosystem projects sit on the Claude Code CLI, which the SDK wraps, rather than the Python/TS SDK API directly.
  • The famous agents that aren't built on any of these SDKs — OpenHands, Dify, RAGFlow, BrowserUse, OpenManus — are the norm, not the exception: the framework layer is not a prerequisite for a successful agent product.
  • Private enterprise deployments (Klarna, Box, etc.) are cited from vendor/public statements, not verified code.

The pattern: public repos tell you who can build fast (solo devs skew toward best-DX frameworks); private enterprise names tell you who will pay (LangGraph wins there).

Bottom line

On popularity: OpenAI Agents SDK beats Pydantic AI on stars, momentum, single-package downloads, and JS reach; Pydantic AI wins trust rankings and has the larger combined Python install surface once pydantic-ai-slim counts. Both trail LangGraph in production-download volume, CrewAI in stars among pure frameworks, Vercel AI SDK in npm reach, and BrowserUse in momentum.

On philosophy: pick by defaults, not popularity — minimal loop on OpenAI → Agents SDK; typed/validated Python → Pydantic AI; state machines with approval gates → LangGraph; role-shaped teams → CrewAI; coding agents on Claude → Claude Agent SDK; TS portability → Vercel AI SDK (+ Mastra for orchestration). Most production teams mix all three camps.

On downstream: the strongest adoption evidence is a mix of lab reference apps, branded company tools (Coinbase), viral community projects (TradingAgents, claude-mem), and private enterprise deployments. Public repos measure developer velocity; enterprise names measure willingness to pay.

Sources