We built agent-express around the idea that most agent failures aren't about the LLM they're about everything around it: cost runaway, context overflow, silent tool errors, no retry logic. Curious what's been the biggest pain point for others building agents for real workloads?
Most agent frameworks invent graphs, chains, and orchestration layers. But when you break it down retry, budget caps, memory, logging each one is just a (ctx, next) function. Same pattern web developers have used for 15 years. Where do you think the middleware model breaks down for agents?
Agent Express is an open-source middleware framework for building AI agents in TypeScript.
If you've written Express.js middleware, you already know how to build agents. The model-tool-model loop in an agent is structurally identical to the request-response cycle in a web server — a context flows through composable (ctx, next) functions.
Three concepts: Agent, Session, Middleware. Five hooks. Built-in middlewares for budget tracking, guardrails, observability, memory compaction and others.
