Self-initiated concept · Human-in-the-loop · 2026
Relay
An agentic operations console where the human gate is the interface — my production agent-platform patterns, re-expressed clean-room in a neutral logistics world.
Concept · Playable belowThe problem
Multi-agent AI fails in trust before it fails in output: pipelines act invisibly, and 'human oversight' is usually a checkbox bolted on after the fact.
What I did
Designed the oversight in as the centrepiece — six specialist agents with visible handoffs, and an approval gate that carries evidence, price and reversibility. Built as a playable, deterministic React component.
Where it landed
A running concept where you are the gate: approve the reroute or reject it, and the pipeline genuinely resolves differently. The pattern comes from my production platform work; the code and data here are all new.
Be the gate
A shipment misses its transshipment. Six agents pick the case up in turn — then the pipeline stops on a proposal and waits for you. There's no timer and no default: it waits as long as you do.
Deterministic demo — one scripted scenario, fictional shipper and figures, no RNG, no network calls. Replay it and answer the other way; the outcome is designed to differ.
Where this comes from
At Ensign InfoSecurity I own the product design, UI/UX and frontend of an agentic AI platform exhibited at GovWare 2025 and 2026 — work that serves government clients and stays confidential. Relay re-expresses that platform's core interaction patterns — specialist agents, visible handoffs, and above all the human approval gate — in a deliberately neutral logistics world: all-new code, fictional data, nothing of my employer's in it. The patterns are my craft to carry; the artifacts are not.
The production platform itself is available to walk through in an interview.
Overview
One case, six agents, one human. The console has three surfaces: a pipeline lane that shows who holds the case at every moment, a transcript where each agent's work lands as it happens, and the approval card — the moment the machine stops and the interface becomes a question. The visitor's answer is real: approve and the reroute books, decline and the fallback plays out, each with its own comms, its own ledger entry, and its own report.
My role
Solo: the scenario design, the interaction and state design, the visual system, and the React/TypeScript implementation on this page. It is deliberately the same working style as the production platform this echoes — designer and engineer being the same person is why the gate behaves exactly as specified.
The design decisions behind it
Everything follows from one conviction: in agentic systems, the human moment is the product. The rest of the interface exists to earn that moment's trust.
Proposals carry evidence, price and reversibility
The approval card is a fixed anatomy: what the agents found, what the action costs, and whether it can be undone. A gate that asks for a decision without those three is asking for a rubber stamp — and rubber stamps are how oversight dies.
Handoffs are visible, always
The lane shows who holds the case at every second — working, done, waiting. Multi-agent systems lose trust in their opacity long before they lose it in their answers; showing the relay is what makes six agents feel like a team instead of a black box.
The human sits in the lane, not above it
The gate renders as a seventh teammate with the same chip anatomy as the agents — and the pipeline waits indefinitely. No countdown, no default action, no auto-approve. Urgency is the scenario's problem; pressure is never the interface's tool.
Rejection is designed as carefully as approval
Decline isn't a dead end — it triggers its own full path: fallback plan, customer comms, audit entry. A gate where 'no' leads nowhere trains operators to always say 'yes', which is the quiet failure mode of human-in-the-loop design.
Status & honesty
Relay is a concept and says so on the console: one scripted scenario, fictional names and figures, no live systems. What's real is the pattern language — evidence-first approvals, visible handoffs, first-class rejection — and the fact that it runs, designed and implemented by one person inside this page.
Next case study: CyberGuard