13 Frameworks and SDKs for Building AI Agents (Turing Post)
Alyona Vert for turing-post, 2026-08-09. A survey of the layer this spoke has mostly stepped over: not the harness you run or the skillpack you install, but the library you build an agent with. Its premise — “Building an AI agent today rarely means starting from scratch” — is the whole reason the agent-framework layer exists.
The thirteen
| Framework | Languages | What it is for | Stated limit |
|---|---|---|---|
| OpenAI Agents SDK | Python, TS | lightweight runtime: tools, handoffs, guardrails, memory, approval, voice | OpenAI-leaning despite provider-agnostic claim |
| LangGraph | Python, TS | low-level orchestration of long-running stateful agents; 39k★ | complexity, steep curve |
| Google ADK | Python | graph workflows, nesting, structured delegation | Python-only, younger ecosystem |
| Microsoft Agent Framework | Python, .NET | enterprise: A2A coordination, checkpointing | heavier |
| Pydantic AI | Python | type-safe agents, dependency injection, structured output | niche to typing-focused teams |
| CrewAI | Python | role-based multi-agent crews and tasks | abstraction overhead when simple |
| Agno | Python | a whole platform: storage, scheduling, access control | opinionated, lock-in risk |
| Strands Agents SDK | Python, TS | model-driven, provider-agnostic with a Bedrock default | small community |
| Mastra | TypeScript | TS-native: graphs, memory, RAG, MCP, Next.js | JS ecosystem only |
| smolagents | Python | minimal; CodeAgent writes Python, sandboxed | little hand-holding |
| LlamaIndex / LlamaAgents | Python, TS | RAG and data: documents, PDF, SQL | specialized |
| CAMEL-AI | Python | role-play, workforce orchestration, large simulations | research-grade |
| LiveKit Agents | Python, TS | realtime voice and multimodal: WebRTC, telephony, turn detection | not for text-only |
Its verdict is a refusal to pick: “There is no single best option.” Choice follows the application and the stack.
Two ladders that both call the middle rung “SDK”
The article’s framing is AI model → API → SDK → agent framework, where an SDK is “a ready-made toolbox” and a framework adds orchestration, memory, tool use, workflows and multi-agent coordination. That ladder measures how much is pre-built for you.
This wiki already holds a different one. claude-agent-sdk sits on Anthropic’s Client SDK → Agent SDK → Managed Agents ladder, which measures who runs the loop and where — you, the library in your process, or the vendor’s sandbox. Both put “SDK” in the middle and mean something else by it. Worth keeping straight, because a source that says “just an SDK” could be making either claim.
What the list says by its shape
Python is the language of this layer. Twelve of the thirteen support it and Mastra is the only one that does not. TypeScript appears in six. Nothing else appears at all, and .NET only inside Microsoft’s entry. For a spoke whose harness corpus is heavily TypeScript and shell, that is a sharper divide than anything previously recorded here.
The layer has started specializing by use case. Voice and telephony (LiveKit), documents and RAG (LlamaIndex), simulation and research (CAMEL), enterprise coordination (Microsoft). A market where the general-purpose entrants have stopped being the only entrants is a market past its first phase.
Anthropic is not on the list. Three of the four large vendors are — OpenAI, Google, Microsoft — plus AWS through Strands’ Bedrock default. The claude-agent-sdk this wiki pages as “the clearest artifact of the model-provider → agent-platform shift” does not appear. This is an observation about the article’s selection and nothing more; the survey states no inclusion criteria, so there is no way to tell an omission from a judgment.
Seen twice, 2026-08-10. llm-engineer-toolkit, a 150+-entry catalog published a day later and shaped quite differently, also carries no Anthropic agent library — the word occurs once in the whole list, inside a monitoring tool’s description of what it instruments. Two criteria-free lists is still not evidence about the market, but it is no longer one author’s oversight.
Checked against a second census
llm-engineer-toolkit names 27 agent libraries to this article’s 13, and seven appear on both. The six of these thirteen it omits are the vendor SDKs (ADK, Microsoft Agent Framework, Strands), the non-Python entries (Mastra, LiveKit), and LlamaIndex — which that catalog files under application development instead. Running the other way, this article has none of the community layer (AutoGen, Swarms, Composio, Browser Use, AgentOps, Langroid) the catalog is full of.
Read together: this list is shaped like the vendor landscape, the other like the Python open-source community, and neither says so. See agent-framework.
What it does not do
No evaluation. The “weaknesses” column is editorial judgment — “abstraction overhead,” “steeper learning curve,” “vendor lock-in risk” — not measurement, and the only number in the piece is LangGraph’s star count. That keeps it consistent with this spoke’s standing finding that adoption and measurement are uncorrelated in this corner, and it means the comparison axes are usable for orientation and not for a decision.
No sourcing for the claims about each framework either; they read as distilled documentation.
Tier
T3. A newsletter survey: secondary, unmeasured, no inclusion criteria stated, and no author
affiliation given beyond the publication. Its value is the census and the axes, not the verdicts.
freshness: volatile — a snapshot of a layer where star counts and feature sets move monthly.
Related
agent-framework · claude-agent-sdk · adk · langchain · a2a-protocol · model-context-protocol · agentic-coding-harness · turing-post