Frameworks & tools
This is the perishable companion to Choosing a framework, which owns the durable decision logic: when a framework earns its keep, when the raw SDK wins. This page is the current landscape—what's shipping, who's mature, who's in maintenance mode. Reviewed quarterly; every star count and version carries an as-of date.
Cards are grouped by job-to-be-done, not vendor. Within each group, the first card is the one we'd reach for by default; the rest are the cases where it's wrong.
Agent & workflow frameworks
LangGraph
Low-level stateful agent runtime from LangChain Inc. Execution modeled as a directed graph: nodes are Python functions, edges are routing rules, a checkpointer serializes state after every node so the run survives a process restart.
| Latest | langgraph==1.2.4 (June 2, 2026) |
| Stars | 34,458 (as of June 2026) |
| Pick it when | Long-running, interruptible workflows; agents with complex conditional routing; teams already on LangSmith for tracing. |
| Skip it when | Your workflow fits in a single function, or you need provider-agnostic tracing (LangGraph couples tightly to LangSmith). |
| Production users | Uber, LinkedIn, Klarna (named in the v1.0 announcement). |
| Docs | docs.langchain.com |
PydanticAI
Type-safe agent framework from the Pydantic team. Agents are generic over result type and dependency type; the framework auto-retries tool calls when the LLM returns invalid data, using Pydantic validation directly.
| Latest | v1.107.0 (June 10, 2026) |
| Stars | 17,700 (as of June 2026) |
| Pick it when | Teams already using Pydantic; single-agent or light multi-agent setups; apps where structured-output reliability is primary. |
| Skip it when | You need durable checkpointing out of the box (no PostgresSaver equivalent), or you need the largest community ecosystem. |
| Docs | ai.pydantic.dev |
CrewAI
Standalone multi-agent framework (no LangChain dependency since v1.0). Agents modeled as role-playing participants with role/goal/backstory. Two primitives: Crews (autonomous collaboration) and Flows (event-driven pipelines).
| Latest | 1.14.7 (June 11, 2026); 201 releases, weekly cadence |
| Stars | 53,300 (as of June 2026) |
| Pick it when | Role-based multi-agent workflows where personas matter; event-driven pipelines with conditional branching. |
| Skip it when | You need fine-grained durable state, or the task is simple enough that role/goal/backstory is just prompt overhead. |
| Docs | docs.crewai.com |
AutoGen (Microsoft)
Multi-agent conversation framework. Original repo (microsoft/autogen, 58,900 stars) is now in maintenance mode as of early 2026. Active development moved to microsoft/agent-framework. The new framework drops the conversation metaphor for event-driven orchestration.
| Status | AutoGen 0.4 — maintenance mode. New: Agent Framework (preview) |
| Stars | 58,900 (legacy repo, as of June 2026) |
| Pick it when | Only if you have an existing AutoGen deployment. For greenfield, evaluate Agent Framework or LangGraph instead. |
| Skip it when | Starting fresh (maintenance mode means declining community investment). |
| Docs | microsoft.github.io/autogen |
DSPy
Prompt optimization framework from Stanford NLP. You write programs with typed signatures; DSPy compiles them into optimized prompts/few-shot sets via teleprompters (optimizers) that search over prompt strategies.
| Latest | 3.2.1 (June 2026) |
| Stars | 35,000 (as of June 2026) |
| Pick it when | You have a labeled eval set and want automatic prompt/few-shot optimization; pipelines with multiple cascading LLM calls that need joint tuning. |
| Skip it when | You don't have eval data to optimize against, or your system has one simple prompt that doesn't justify the abstraction overhead. |
| Docs | dspy.ai |
RAG & data frameworks
LlamaIndex
Data framework for LLM applications. Owns the indexing-to-query pipeline: document loaders, chunking strategies, vector store integrations, query engines, and retrieval abstractions. The most mature RAG-specific toolkit.
| Latest | llama-index-core==0.14.22 (June 2026) |
| Stars | 50,100 (as of June 2026) |
| Pick it when | Complex document ingestion pipelines; multi-source RAG with metadata routing; you need pre-built integrations for 160+ data sources. |
| Skip it when | Simple single-source retrieval where pgvector + 50 lines of code does the job, or you want to avoid the large dependency tree. |
| Docs | docs.llamaindex.ai |
Infrastructure
LiteLLM
Unified API gateway for 100+ LLM providers. One completion() call, swap providers by changing the model string. Adds load balancing, fallbacks, spend tracking, and rate limiting across all providers.
| Latest | v1.88.1 (June 2026) |
| Stars | 50,200 (as of June 2026) |
| Pick it when | Multi-provider routing, fallbacks, or spend tracking; you want OpenAI-compatible interface to Anthropic/Gemini/open-weight models. |
| Skip it when | Single-provider app with no routing needs; the extra abstraction layer adds latency (~2-5ms) you can't afford on voice paths. |
| Docs | docs.litellm.ai |
vLLM
High-throughput LLM serving engine. PagedAttention for efficient KV cache management, continuous batching, tensor/pipeline parallelism. The default self-hosting inference server in 2026.
| Latest | v0.8.x (June 2026) |
| Stars | ~50,000 (as of June 2026) |
| Pick it when | Self-hosting open-weight models; you need continuous batching, speculative decoding, or multi-GPU inference. |
| Skip it when | Using hosted APIs exclusively; your traffic is too low to justify GPU ops overhead. |
| Docs | docs.vllm.ai |
Voice & realtime
LiveKit Agents
Real-time voice/video AI agent framework built on the LiveKit WebRTC infrastructure. Handles VAD, STT, LLM orchestration, TTS, and turn-taking as a pipeline with pluggable components.
| Latest | v1.6.x (mid-2026) |
| Stars | ~4,200 (agents repo, as of early 2026) |
| Pick it when | Production voice agents with WebRTC transport; you need infrastructure (TURN, rooms, recording) alongside the AI pipeline. |
| Skip it when | Text-only agents, or you're using a native speech-to-speech model (OpenAI Realtime API) that bypasses the cascaded pipeline entirely. |
| Docs | docs.livekit.io/agents |
Pipecat
Open-source framework for building voice and multimodal AI agents. Frame-based pipeline architecture: audio/video frames flow through processors (VAD, STT, LLM, TTS) that you compose declaratively.
| Latest | v0.0.55 (early 2026) |
| Stars | ~12,000 (as of early 2026) |
| Pick it when | Multimodal agents (voice + vision); you want a lighter, more composable alternative to LiveKit's full infrastructure stack. |
| Skip it when | You need production-grade WebRTC rooms, recording, and TURN servers out of the box (LiveKit owns that layer). |
| Docs | docs.pipecat.ai |
Observability
Langfuse
Open-source LLM observability platform. Tracing, prompt management, evaluation, and cost tracking with full OpenTelemetry support. Self-hostable or cloud.
| Latest | v3.185.0 (June 2026) |
| Stars | 29,000 (as of June 2026) |
| Pick it when | You want open-source observability you can self-host; OTel-native tracing matters; budget-conscious teams. |
| Skip it when | Your team is already deep in LangSmith and unwilling to migrate; you need enterprise SSO/RBAC without self-hosting. |
| Docs | langfuse.com/docs |
LangSmith
Closed-source observability and evaluation platform from LangChain Inc. Tightly integrated with LangGraph and the LangChain ecosystem. Strongest eval/annotation tooling for teams in that ecosystem.
| Latest | Cloud SaaS (continuously deployed) |
| Pick it when | You're already using LangGraph/LangChain; you want the tightest integration with those tools; enterprise procurement is not an issue. |
| Skip it when | You need self-hosting, OTel-native export, or provider-agnostic tooling. Vendor lock-in is real. |
| Docs | docs.smith.langchain.com |
Phoenix (Arize)
Open-source LLM observability tool focused on tracing, evaluation, and experimentation. OTel-native with OpenInference conventions. Lightweight alternative to the full-platform plays.
| Latest | v10.x (June 2026) |
| Stars | 10,100 (as of June 2026) |
| Pick it when | Local-first development; you want fast trace visualization without standing up infrastructure; already using Arize for ML monitoring. |
| Skip it when | You need a full production platform with prompt management, annotation queues, and team collaboration (Langfuse or LangSmith). Note: ELv2 license, not true open-source. |
| Docs | docs.arize.com/phoenix |
Browser automation
browser-use
AI-native browser agent framework. Connects LLMs to a browser via DOM extraction and structured actions. The model sees a cleaned DOM tree and issues high-level commands (click, type, navigate).
| Latest | v0.13.1 (June 2026) |
| Stars | 98,400 (as of June 2026) |
| Pick it when | Building AI agents that need to browse the web; you want DOM-driven actions (faster, cheaper than screenshots). |
| Skip it when | The target site is behind auth you control (use their API instead), or you need screenshot-driven universality (use Computer Use). |
| Docs | docs.browser-use.com |
Playwright
Cross-browser automation library from Microsoft. Not AI-specific, but the default execution substrate under AI browser agents. Headless Chromium/Firefox/WebKit with auto-waiting, network interception, and codegen.
| Latest | v1.52.x (early 2026) |
| Stars | ~89,800 (as of early 2026) |
| Pick it when | You need the browser execution layer under browser-use or your own agent; E2E testing of AI-powered UIs; web scraping at scale. |
| Skip it when | You're using Computer Use (screenshot-driven, no DOM access needed), or the task doesn't involve a browser at all. |
| Docs | playwright.dev/python |
Last reviewed: June 12, 2026. Star counts and versions shift weekly. If a number here is more than 90 days old, check the linked docs. The decision logic for when to use a framework at all lives in Chapter 7.12.