Public demo

Agent guardrail designer

Pick a task, then decide what the agent is permitted to do. You get a risk level, the approvals it requires, a plain-language rules file and a simulated session that stops for a person.

This is a design exercise styled like a terminal coding agent. Nothing here calls an AI, and nothing you choose is sent anywhere.

> /task Draft Jira stories from an idea Summarize meeting notes Answer questions from a document library

> /permissions Read look at the material for the task Draft write something for a person to review Create add new items where the team works Edit change items that already exist Prepare outside messages for people outside the team, never sent by the agent

> risk

> approvals (the system must enforce these)

> cat ground-rules.md

These rules are requests written into the agent's instructions. A request can be ignored, so approvals, permissions and the action log must be enforced by the system around the agent, not only requested.

> run agent (eight steps, and it stops wherever a person must approve)