03 · Software patterns

Beyond the Interface.

Today's platforms are interface-heavy because humans did all the work and needed screens to do it. With capable agents, that premise inverts. The future platform has almost no standing UI — the agent performs the workflow, and an interface materialises only at the moment a human judgement is genuinely required, then dissolves.

Today · dashboard
Post-UI · agent-native
agent at work
Approve £180 goodwill credit for customer #4821?
Approve Decline
Left — a dashboard you must read and operate. Right — an empty surface; a single decision card surfaces only when judgement is required, then vanishes.

The number that matters

The point is a step-change in human effort, not a marginal UX improvement. The target is roughly a 90% reduction in human touchpoints: agents own the routine, predictable 90%; humans handle the remaining 10%. Stating the number is essential — it prevents the failure mode of building agentic plumbing while keeping every human in their old seat.

90% · AGENT
10%
routine, predictable, repeatablegenuine judgement

Defining the 10% — the crux

The reserved 10% is genuine human judgement — risk, strategy, ethics, ambiguity with real stakes — not "the agent got stuck." A stuck agent is an engineering defect to be learned away, not a permanent human task. The research contribution is partly how you tell the two apart, and how the system shrinks "stuck" toward zero while protecting "judgement."

What it looks like in practice

  • No dashboards to log into. The agent runs the process end-to-end.
  • At a true decision point, a micro-UI surfaces — a single decision, the minimal data needed to make it, nothing more — then disappears.
  • Inputs can be batched (a queue of judgement calls) or real-time (an urgent approval).

Concrete examples

01

Customer operations

Agent resolves the routine ticket. A one-tap judgement card appears only for goodwill or refund exceptions.

02

Internal back-office

Procurement, onboarding, reconciliation run headless. Humans approve only policy-edge cases.

03

Content & knowledge

The "website" stops being a destination you browse. The agent retrieves and acts, surfacing a data point only where you must choose.

The scope is effectively unlimited — wherever a screen today exists mainly to let a human do repetitive work.

Why now

Enterprises are spending heavily on agentic AI but mostly bolting agents onto UI-first products, capturing a fraction of the value. A clear standard for agent-native, UI-minimal design is the missing primitive.

Next · 04 — The Self-Discovering Data Mesh