Pydantic opens Agent Sandboxes to TypeScript apps with a Vercel AI SDK example
samuelcolvin · x · 2026-11-16
Pydantic announced its Agent Sandboxes now work with non-Python apps: the sandbox code stays Python while the client can be Python or TypeScript, with a cloneable Vercel AI SDK example in pydantic/logfire-js.
The example has an agent review a company's Q3 travel spending by writing Python that calls tools in one batch instead of one model turn at a time (4 departments, 24 employees, 368 multi-currency line items).
Key patterns covered:
- Large data, small answer: line items stay in the sandbox; the model only receives totals and 8 over-budget employees
- Parallel calls via asyncio.gather
- Error and recovery: a missing CHF rate error guides the model to finance's fixed rate
- 50,000 free sandbox runs per month
More from coding & agent
- One prompt, no SaaS: rebuilding email and inbox tools as a Claude artifact — katieruthmishra · 2026-11-16
- Top AI engineer: reliable agents need boring deterministic steps, not more LLMs — omarsar0 · 2026-11-16
- ChatGPT sets up remote PC via custom MCP while Grok bot books flights — cgarciae88 · 2026-11-16
- Perplexity: Median Employee Now Gets 5.4 Agent-Hours a Week, Up 11x in 10 Weeks — _andychung · 2026-11-16
- Old world: write code, then test. New world: build the verifier, then let agents grind — Yuchenj_UW · 2026-11-16
- Claude Code + Raspberry Pi Homebridge: a Raspberry Pi smart home stack ripe for automation — mobileraj · 2026-11-16