Free AI template
Agent flow with guardrails
An order-status agent with a step limit, input and output guardrails and read-only tools, that passes the validator.
It opens with this template loaded and runs in your browser. You don't need an account to use it; a free account saves your work and exports it.
The job it does
When you design an agent: check that loops end, risky tools sit behind a person and every path ends in a reply. It comes with its own checks, so after you change it you can see at a glance that it still does the job.
What's inside
This is the template exactly as the AI Lab opens it, read-only. Open it to run it, change it and export it.
- 6 steps
- 6 arrows
An agent that answers "where is my order?": a guardrail on the way in, a model step with a step limit, a read-only tool for the customer's own order, and a guardrail before the reply. It passes the validator with no warnings. Add your own tools and keep it green.
- The agent flow passes the validator with no warnings
| Step | Kind | Settings |
|---|---|---|
| Customer message | User message | |
| Guardrail (in) | Guardrail | checks the input; stops on: ignore previous, system:, card number, password |
| Answer about the order | Model step | Find the order in the customer's own account and say where it is. Never guess a time.; at most 3 steps |
| Look up my order | Tool | read only |
| Guardrail (out) | Guardrail | checks the output; stops on: refund, voucher, phone number, address |
| Reply | Reply |
- Customer message → Guardrail (in)
- Guardrail (in) → Answer about the order
- Answer about the order → Look up my order
- Look up my order → Answer about the order
- Answer about the order → Guardrail (out)
- Guardrail (out) → Reply
The lesson behind it
Learn why it works, with a podcast and a quiz for each lesson.
More AI templates
Cost estimate for an AI feature
A prompt, the expected reply length and traffic, and what it costs per call, per day and per month on each model.
Prompt review
A support prompt in parts (role, task, context, rules, format, an example) that passes the linter, with its token count and cost.
Eval set for a chatbot
Replies a bot really gave, labelled good or bad, and checks that agree with every label, exportable as CSV.