The Routecraft Blog
Tutorials, postmortems, and small notes from building Routecraft in production.
System prompts are requests, not rules. Human review, LLM judges, and guardrail frameworks all raise the cost of an attack, and none of them is a boundary. Banks did not solve embezzlement by hiring more honest tellers. Give the agent hands, not keys.
The personal-assistant wave gave every employee an agent with their own access, their own memory, and their own chat window. That model cannot serve a team. The unit of organisational AI leverage is the governed capability, not the personal agent.
TypeSafe AI's Jev answers typed questions with probabilities instead of text, which makes it a fast first judge for agent results and a reason to call the slow one far less often. Here is the route, the three things to know before you call the SDK, and what Jev is built for beyond judging.
A chat loop with tools is not a harness. An agent that works for a team needs four primitives most personal harnesses skip entirely: delegation, a learning loop, self-aware capability gaps, and multi-channel presence, plus three platform rules that keep the whole thing governable.
Every organisation is building a central repo of AI skills, prompts, and agent definitions right now. It is the right first step, and it has a ceiling nobody talks about. A skill is advice for an AI assistant; the doing still runs on personal access. A maturity ladder for what comes next.
A ten-minute walkthrough from `bunx create-routecraft` to a working MCP server that Claude Desktop and Cursor can call. No auth, no infrastructure, no boilerplate. Just a TypeScript function exposed to your AI agent.
Build with Routecraft
Write capabilities, expose them as MCP tools, and let agents do real work on your terms.