2026-10-25 · America/Los_Angeles · 社区动态 · #8

I built a small validation layer that stops LLM agents from sending bad tool calls (TS + Python, MIT

Agents call tools with arguments the model made up: an order ID that doesn’t exist, a negative amount, a date like “yesterday”. Most of us patch this with try/except inside every tool. agentic-gate sits between the model and your tools. It validates the arguments against a Zod (or Pydantic) schema before anything runs, sends the validation error back to the model so it can fix the call, and trips a circuit breaker if the same call keeps failing, so you don’t burn 50 turns in a retry loop. It’s provider-agnostic: adapters for OpenAI, Anthropic, Bedrock and Gemini, plus an async validate()…

热度 46.0 / 100；排名与评分保留该期记录。