The problem
AI agents are starting to act, not just answer: sending payments, emailing customers, calling APIs. Every one of those actions deserves the same thing a junior employee would get: a quick human look before it goes out. Today most agent frameworks log what happened, but nothing gives the operator a calm place to decide what should happen.
The idea
Gate Console is the human side of an agent-safety gate. When an agent wants to perform a consequential action, the gate holds it in an approval queue and shows the evidence a person needs to decide in seconds, not minutes.
What the prototype does
- Approval queue: pending agent tool calls with a verdict pill on each (block / review / pass), so the operator sees risk at a glance.
- Decision detail: the screen a reviewer actually works in. It shows why the gate held a payment: the injected instruction found inside a retrieved invoice (quoted as untrusted data, visually separated from the agent's own plan), the policy that fired, and clear Approve / Block / Escalate paths with what each choice means.
- Audit log: every gate decision, immutable, newest first.
- Policy settings: the knobs behind the gate (wire ceiling, payee cooling period, egress allow-list) as plain-language switches.
What inspired it
Building agent prototypes in past hackathons kept hitting the same wall: the scariest part of agentic AI is not the model, it is the moment an automated action leaves the building. Prompt-injection attacks arrive inside ordinary-looking documents, so a reviewer needs to see the hostile text next to the agent's plan, not buried in a log.
Design decisions
- Dark, high-contrast theme tuned for long review sessions, with AA-contrast text.
- Evidence is always labeled as untrusted data and quoted, never paraphrased.
- Verdict pills pair color with a word, so color alone never carries the meaning (accessibility).
- Rows and toggles are keyboard reachable; every screen has visible focus states.
- One decision per screen. The reviewer's next action is always the most prominent button.
How it was built
A static front-end prototype: HTML, CSS and vanilla JavaScript, no backend needed. Demo data is synthetic; no real payments, inboxes, or agents are involved. It runs at the Try-it-out link.
Challenges
The hard part was not drawing a dashboard, it was deciding what to leave off. Early drafts crammed charts in; the final design gives every screen one job, because a reviewer under pressure does not read, they scan.
Log in or sign up for Devpost to join the conversation.