Demo
An agent that asks before it acts
Chat with a pretend project agent about an invented team’s sprint board. Small requests run at once, changes wait for you to confirm, and anything that cannot be undone waits for an approval that expires and works only once.
This is a simulation. No AI model runs on this page: the agent’s replies are scripted, and the company, people, tasks and numbers are all invented. Nothing you type leaves your browser or is kept, and reloading the page starts again.
Conversation
You are talking to the project agent of Driftwood, an invented app studio. Every request goes through a gate before anything runs, and the activity log shows each step.
Low risk: runs at once
Medium risk: asks you to confirm
High risk: needs an approval
Costly: uses most of the budget
Budget and time limit
Demo budget
0 of 20 credits used. Running.
Each request costs a credit for the model’s turn and one for each tool call, the way a real model costs money. When the budget is spent, a kill switch stops the agent until a person resets it.
Choose the longer time if you need more time to read. Each approval keeps the time it was given.
Activity log
Each step will appear here: the request, the gate’s decision, any confirmation or approval, the action and its result.
Sprint board
Day 6 of 10. 11 of 30 points done (37%).
- T-1Watering reminder screenDone
- T-2Plant photo uploadIn progress
- T-3Search by plant nameDone
- T-4Dark theme coloursIn review
- T-5Fix the crash on an empty plant listDone
- T-6Offline plant listTo do
- T-7Remove the old export buttonTo do
- T-8Care tips pageIn progress
Hazel leads the team, and Robin, Wren, Basil and Juniper are its engineers. They are building a plant care app.
What the demo shows
- A gate with three tiers. Every tool the agent can use declares how risky it is. Reading the board runs at once; a change that can be undone asks you first; a change that cannot be undone becomes an approval card.
- Approvals that cannot be replayed. An approval runs exactly the action stored when it was proposed, only once, and only before it expires. A second click, a late click or a click after a denial runs nothing.
- Figures from code. Every number in a reply is worked out by code from the board, and each reply can show how. In a real system the model explains the numbers; it never writes them.
- A budget kill switch. Requests are checked against a spending limit before they run. When it is reached, the agent stops and anything waiting is cancelled until a person resets it.
- A record of every step. The activity log shows each request, the gate’s decision, each confirmation or approval, and what ran.