Create an Agent

Scaffold a trustworthy AI agent · in your browser

A single model call answers a question. An agent is the small program that makes the answer trustworthy: it carries a directive and runs one disciplined loop: sample → evaluate → publish → verify, with a separate verifier whose only job is to say no. Fill in the fields and download a ready-to-use agent spec. Nothing is uploaded; it all runs in this tab.

Agent name Role

Directorcarries the directive; decides the next step

Workerdoes one narrow task (search, summarize, extract)

Verifiersays no when output breaks the rules

Task (the one job this agent does) Directive (paste one, or make it in the Directive wizard) EVALUATE: the core action (what it actually does each turn) VERIFY: the check that can reject it (the falsifiable test) Tools / retrievers it may call (optional, one per line) Model policy

Download agent

The Markdown spec doubles as a system prompt; the JSON is a config your runtime can load. The four-phase loop is filled in from your answers.

About this project

This wizard is a free portfolio piece showing how I build agents: a directive plus one disciplined loop and a verifier that can reject. It runs entirely client-side. I'm seeking a full-time AI engineering role.

Contact for hiring ← Create a Directive How my agents work