AI Agent Frameworks

This is the layer above raw LLM APIs (see AI / LLM Tooling). Agent frameworks add tool calling, state, planning, retries, evals hooks, and (often) durability.

TypeScript-first

• ★ Mastra — TS-first; agents, workflows, RAG, evals, voice, observability. The default for new TS agent projects in 2026.
• ★ Vercel AI SDK + AI SDK Agents — same SDK as the chat hooks; first-class agent loop with tool calling, generative UI streaming, and multi-step reasoning.
• OpenAI Agents SDK — official from OpenAI; TS + Python; clean API; works well with Responses API and agents-as-tools patterns.
• Anthropic Agent SDK — same shape from Anthropic; TS + Python; strong tool-use defaults.
• LangGraph + LangGraphJS — graph-based agent orchestration; the most flexible for branching / cyclic agent flows.
• LangChain.js — broader framework; mature; somewhat heavy.
• LlamaIndex.ts — RAG + agents; pairs well for knowledge-base agents.
• Inngest Agents — durable agent loops on top of Inngest's workflow engine; production-grade retries / timeouts.

Python-first (worth knowing if you do hybrid stacks)

• Pydantic AI — Pydantic's framework; clean types.
• smolagents (Hugging Face) — code-first agents; minimal abstractions.
• LangGraph — most popular Python agent orchestration.
• CrewAI — multi-agent collaboration.
• AutoGen (Microsoft) — multi-agent.
• DSPy — declarative; treats prompts as compiled artifacts.

Browser / web automation agents

• ★ Stagehand (Browserbase) — natural-language Playwright; the popular choice for "agent that uses a browser." See Web Scraping.
• browser-use — Python competitor; LangChain-flavored.
• Skyvern — open-source RPA-flavored.
• Playwright + your own LLM loop — DIY when frameworks feel like too much.

Tool registries / runtimes

• See MCP for the standard protocol agents use to call external tools.
• OpenAI Functions / Anthropic Tools — provider-native tool calling; works without a framework.
• @vercel/ai tool helper — typed tool defs with Zod.
• Mastra Tools — typed tools registered to agents.

Memory / state

• @upstash/vector, pgvector, Pinecone, Convex / Vectorize — vector stores for long-term memory.
• mem0 — open-source agent memory layer.
• Letta (formerly MemGPT) — agent memory + long context.
• Redis + simple key-value — often enough for session memory.

Multi-agent orchestration patterns

• Supervisor / orchestrator — one agent dispatches to specialists.
• Sequential — chain of single-purpose agents.
• Parallel + voting — N agents, take majority.
• Reflective — agent critiques its own output before returning.

Production concerns

• Determinism / replay — reach for Workflow Engines (Inngest, Trigger.dev, Temporal, Restate).
• Observability — AI Evals and tracing (LangSmith, Langfuse, Braintrust, Helicone).
• Cost — AI Gateways for caching + routing.
• Safety — Content Moderation on both inputs and outputs.

Pick this if…

• Default new TS agent project: Mastra or Vercel AI SDK Agents.
• Sticking with one provider: OpenAI Agents SDK or Anthropic Agent SDK.
• Complex branching / cyclic flows: LangGraphJS.
• Durable agents that must survive restarts: Inngest Agents or Trigger.dev.
• Browser-using agent: Stagehand.
• RAG-first knowledge agent: LlamaIndex.ts or Mastra RAG.