Most TypeScript teams shopping for an agent framework don't need one. A single generateObject call covers classification, extraction, summarization, tagging — the 80% case for production LLM work in TS right now.
But once the model starts deciding what to do next, surviving deploys, or coordinating with other agents, you start shopping. And the moment you do, you discover the TS agent ecosystem is not one market but three.
Three layers
Loop — wraps the agent loop so you don't write while (!done) {…} yourself. "Let the model decide what to do next."
Mastra, Vercel AI SDK, Claude Agent SDK, OpenAI Agents SDK, LangGraph, AgentKit, VoltAgent, Google ADK.
Runtime — keeps the loop alive across crashes, deploys, and long pauses. "Keep the work alive when servers die."
Mastra Platform, LangGraph Platform, Claude Managed Agents, Inngest, Flue.
Sandbox — runs an existing coding agent in isolation, often in parallel. "Run ten Claude Codes in ten boxes."
Sandcastle, Flue (Daytona mode), E2B, Daytona.
The decision tree
- Building an agent loop? → A loop framework. Mastra (batteries-included) or Vercel AI SDK with Agent (minimal new surface area) is the realistic shortlist.
- Got a loop, need it to survive production? → A runtime. Use the matched host if you've committed to a framework (Mastra Platform for Mastra, LangGraph Platform for LangGraph), otherwise Inngest or Flue.
- Got a coding agent, want to fan it out against a repo? → A sandbox harness. Sandcastle is the cleanest TS-native fit.
Worked example
Support agent that looks up orders, issues refunds, pauses for manager approval over $500.
- Loop complex? No. Mastra or Vercel AI SDK with three tools and a stop condition.
- Needs durable approval pauses? Yes. Inngest or Vercel Workflows.
- Modifying a repo in a sandbox? No. Skip.
One loop + one runtime, no sandbox. Two vendors, not three.
The honest take
Most TS shops reading this don't need any of these yet. The single generateObject call still does an enormous amount of work for an enormous number of teams.
The point isn't go pick a framework. The point is: when the time comes, know what you're picking between. The loop is the agent. The runtime hosts the loop. The sandbox fans out an existing coding agent. Mixing them up is how shops end up paying three vendors that each solve a different problem.
Curious what your team is running, and what bit you in production that you didn't expect. Drop it in the comments.
Full version with per-tool deep dives, code snippets, and picker tables for each layer is on EveryDev.ai: *TypeScript Agent Frameworks in 2026: Loop, Runtime, Sandbox*.
Top comments (0)