Mastra

Mastra is a TypeScript framework for building AI agents, not a pre-built agent you deploy. It comes from the Gatsby team, which shows in the developer experience: clean APIs, good docs, and a Next.js integration that actually works. The core idea is sensible: workflows as first-class primitives, built-in RAG for context retrieval, and human-in-the-loop hooks so you can gate risky actions. I built a customer support triage agent with it that routes tickets, fetches context from our knowledge base, and drafts replies. The workflow system let me chain LLM calls, database lookups, and approval steps without wrestling with state machines or callback hell. The evals framework is genuinely useful: you write test cases, Mastra runs them against your agent, and you get pass/fail metrics. This catches regressions when you tweak prompts. The RAG primitives handle chunking and embedding without forcing you into a specific vector store, which is rare. You bring your own Pinecone or Postgres with pgvector. Where it stumbles: the orchestration is single-threaded, so long-running workflows block. No built-in retry logic for flaky APIs, you wire that yourself. The cloud offering is new and light on features compared to running it yourself. Documentation assumes you know TypeScript well; if you are coming from Python, the learning curve is real. Compared to LangGraph, Mastra is less mature but far easier to reason about. LangGraph gives you more control over execution graphs, Mastra gives you faster iteration. If you are already in the Next.js ecosystem and want agents that feel like application code rather than research prototypes, Mastra is the best option I have found. The human-in-the-loop primitives are particularly good: you can pause a workflow, send a Slack message, wait for approval, then resume. That alone justifies it for production use cases where full autonomy is too risky.