Build Your Own Agent
Everything you have seen in this track comes down to four decisions: what it is for, what it may use, how long it may run, and when it must stop. Make those four decisions here and watch the result run.
Optional: load a real model
1. Who is this agent?
Write what it must always do and what it must never do. Small models follow short, concrete rules far better than long personalities.
2. What may it use?
Everything you leave unchecked is something no instruction and no injection can make it do. This list is a security boundary, not a preference.
3. Run it
The full trace appears here: every thought, tool call, and observation.
4. The same agent, as code
This is the shape of what you just built in any framework. The names differ, the four decisions do not.
Before you take one of these to work. Write five tasks it must get right and turn them into assertions on the testing page. Check every tool you granted against the guardrails checklist. Then run it at assist level for a week and read every trace. That week is what separates the agents that stay in production from the ones quietly switched off a month later.
Free from AI School - no signup, everything runs in your browser.