AI library

Frameworks

Agent and LLM framework landscape

Last reviewed 16 October 2026

Overview

Frameworks provide building blocks for LLM applications: model clients, tool loops, retrieval, memory, orchestration, tracing. They speed up getting started, but add a layer between you and the model that can hide prompts and make debugging harder. Many teams now start from vendor SDKs and add a framework only when it clearly helps.

Key concepts

  • Vendor SDKs and agent kits:
    • Anthropic SDKs (with tool runner) and the Claude Agent SDK (the Claude Code harness as a library), plus Managed Agents for hosted agents.
    • OpenAI SDKs and the OpenAI Agents SDK.
    • Google's Agent Development Kit (ADK).
  • Orchestration frameworks:
    • LangChain / LangGraph: components plus graph-based, stateful agent workflows. LangGraph 1.2 (2026-05-12) resumes error handlers durably after host crashes and adds set_node_defaults(); 1.2.12 added response_schema to interrupt().
    • LlamaIndex: data and retrieval focus, with agent workflows.
    • CrewAI: role-based multi-agent teams ("crews") plus Flows, an event-driven workflow layer. Since 1.15 (2026-06-25) flows and crews can be declared in configuration files, and crewai eval compares models.
    • Microsoft Agent Framework: successor to AutoGen and Semantic Kernel, for .NET and Python. Version 1.0 (2026-04-03) brought stable APIs, long-term support, MCP and A2A interop, and migration guides from both predecessors.
    • Pydantic AI: typed, Python-first agents.
    • Vercel AI SDK: TypeScript toolkit for AI apps, UIs, and agents. Version 7 (2026-06-25) adds per-tool approval policies, durable resumable runs (WorkflowAgent), layered timeouts, a portable sandbox abstraction, OpenTelemetry integration, and an experimental HarnessAgent that wraps Claude Code, Codex, and other harnesses behind one interface.
    • DSPy: programming LLM pipelines with automatic prompt optimisation. Its GEPA optimiser can also change program structure (Flex, since 3.3). 3.4 (2026-09-25) moves model calls to native engines with LiteLLM as fallback and deprecates the legacy LM interfaces for removal in 3.5.
  • Observability and evals: Langfuse, LangSmith, Braintrust, Arize Phoenix, OpenTelemetry GenAI conventions.
  • Interop standards: MCP for tools and data; agent-to-agent protocols (such as A2A) for agents talking to each other.
  • Shared direction (late 2026): the main frameworks are converging on sandboxed code execution, human approval of tool calls, cancellation of long runs, and the newer MCP SDKs.

How to use it

  • Prototype directly with the vendor SDK to see real prompts and responses.
  • Adopt a framework for a specific need: durable state, complex graphs, many providers, built-in tracing.
  • Prefer frameworks that let you see and control the exact prompts.
  • Add tracing from day one, whatever you choose.

Current state

Latest versions checked on 2026-10-04 (GitHub releases pages), except CrewAI and LangGraph, checked later on PyPI because their GitHub release pages show wrong years:

Framework Latest stable Date Note
Microsoft Agent Framework Python 1.20.0 / .NET 1.23.0 2026-10-02 / 2026-10-01 1.0 GA on 2026-04-03
OpenAI Agents SDK (Python) 0.23.1 2026-10-02 Still 0.x
Google ADK (Python) 2.11.0 2026-10-02
Pydantic AI 2.54.0 2026-10-02 Near-daily minor releases
CrewAI 1.15.26 2026-10-08 1.15.0 on 2026-06-25; dates from PyPI (checked 2026-10-09)
LlamaIndex 0.14.25 2026-09-21 Patch release; no note
LangGraph 1.2.14 2026-10-06 1.2.0 on 2026-05-12; dates from PyPI (checked 2026-10-08)
DSPy 3.4.0 2026-09-25 3.0.0 on 2025-08-12; dates from PyPI (checked 2026-10-12)
Vercel AI SDK 7.x 2026-06-25 (7.0) Latest 7.x patch not checked; date from Vercel's announcement (checked 2026-10-16)
  • 2026-10-07: CrewAI 1.15.24 added crewai eval with model swapping via llm_overlay, Oracle integrations, and an experimental job runner.
  • 2026-10-06: LangGraph 1.2.14, the fourteenth patch on the 1.2 line.
  • 2026-10-02: Google ADK 2.11 adds run cancellation, tool approval in workflows, SQLite memory, and opt-in MCP SDK 2.x support.
  • 2026-10-02: OpenAI Agents SDK 0.23 focuses on hardening sandboxes, sessions, MCP, and realtime agents.
  • 2026-10-02: Pydantic AI 2.54; recent 2.5x releases added sandboxed workspaces (E2B, SSH, Bubblewrap, Sprites), plugins, and its own clai2 terminal agent.
  • 2026-09-25: DSPy 3.4, the "LM transition" release: native LM engines, experimental typed decision outputs, and async ReActV2. Legacy interfaces will be removed in 3.5 (no date given).

Open questions

  • Add a primary source for a LlamaIndex major release; it was still on 0.14.x on 2026-10-12. Vercel AI SDK 7 answered on 2026-10-16: Vercel's own announcement confirms the 2026-06-25 release date given by secondary sources.

Sources