Thinking Out Loud · Product & AI

The Interface Is Not the Product

How native agentic environments dismantle a 40-year assumption about software — and what builders need to do about it now.

Topics: Agentic AI · A2A · Product Strategy · UX
Type: Personal thinking piece
01

The assumption nobody questioned

For forty years, software has shipped with a single implicit contract: we build the interface, you use it.

Personalization meant choosing a dark theme. "Customization" meant dragging a column to resize it. The product was the UI, and the UI was the product.

This wasn't laziness. It was the only viable model when humans were the only actors operating software. Designers made deliberate choices about what to show, hide, and enforce — a lossy compression of system capability into something a human could navigate without training.

"The interface was always a lossy compression of what the software could actually do. We just never had a reason to question it."

We now have a reason.

02

Three forces arriving simultaneously

Three distinct shifts are converging, each sufficient on its own to reshape software. Together they make the current model untenable.

The three forces
01
Agentic execution. AI agents can now operate software autonomously. When the actor is not human, visual hierarchy, affordances, and feedback loops serve no one. The "interface" the agent needs is a well-defined capability contract, not a rendered screen.
02
Agent-to-agent protocols (A2A). Agents increasingly talk directly to each other — delegating, composing, negotiating tasks across system boundaries. The interaction model shifts from hub-and-spoke to a mesh. UIX must consolidate at the edges: intent entry and exception handling only.
03
Personal UI agents. Coding agents are now capable enough to let users become their own forward-deployed engineers. The same email client can render as a task manager for one person and an events calendar for another — through user-directed agent customization against a shared primitive layer.
03

What replaces UIX — and what doesn't

UIX doesn't disappear. It bifurcates and, in the near term, explodes in variety before eventually receding.

TodayOne static UI → all users
↓ personal UI agents arrive
Near termPrimitive layer → agent shapes UI per user → human interacts
↓ A2A matures, agents take execution
Further outPrimitive layer → agents operate directly → human sees exceptions only

The old "interface" splits into three distinct surfaces, each serving a different actor:

SurfaceServesWhat it looks like
Capability / intent layerExternal agents (A2A)Tool schemas, intent APIs, agent cards
Semantic primitive layerPersonal UI agentsComposable, semantically described capabilities
Supervisory UIHumansSparse, high-stakes: goal-setting, review, override

Traditional UIX survives — but concentrated into the moments that require genuine human judgment. Think mission control, not cockpit.

04

The new design disciplines

Traditional UX asked: how does a human perceive and move through this? The emerging disciplines ask entirely different questions.

Emerging design disciplines
Agent interface design. How does an agent discover, trust, and safely compose with this capability? Legibility, affordance, and error recovery still matter — expressed in schemas and contracts, not pixels.
Delegation UX. How do humans grant and scope agent authority? Scoped, revocable permissions ("this agent can book rides under $20, never share my location") are the new form design. Trust and authorization become the primary UX problem in A2A environments.
Provenance UI. When Agent A delegated to Agent B on your behalf, can you see the chain? Audit trails become user-facing features, not just compliance logs.
Interrupt design. When an agent mesh escalates a decision to a human, the UI must surface just enough context, just fast enough, to decide and release. Too much and the human is overwhelmed. Too little and they can't make a meaningful choice. A new class of design problem with no prior art.
05

What this means if you're building software today

The instinct is to add an AI chatbot on top of what you already have. That's the wrong move.

For companies with rich UI products — think Uber, Grab, bTaskee, any consumer platform — adding a chatbot on top creates a fragile wrapper over an architecture never designed for non-human actors. The agent ends up scraping UI state or calling internal endpoints never built for external consumption. It breaks constantly and can't be trusted with autonomous actions.

The correct first move is architectural: extract your product's capabilities into a clean intent layer, completely decoupled from the UI. Every meaningful user action becomes a parameterized, idempotent, semantically described API endpoint. The UI becomes one client among many.

"Don't add AI features to your product. Build the layer that makes your product composable — then everything agentic follows."

For operational transformation versus API-layer investment, the sequencing matters. Ops workflow automation delivers ROI today and builds your team's AI muscle in a lower-stakes environment. But build those workflows against clean intent contracts from day one — your internal automations become the prototype of your external capability layer.

The four-phase path

Now
Extract semantic primitives. What your software can DO, expressed cleanly, decoupled from UI. Not a component — a capability description. This is the foundation everything else builds on.
0–6 months
Agentic ops workflows. Internal automation built on primitives. Real ROI, contained risk, genuine AI muscle-building. Each workflow forces you to express a capability cleanly — primitive extraction done organically.
6–18 months
Personal UI agents emerge. Users begin reshaping their experience against the same primitive layer. The user never "builds" anything — they describe, the agent configures, the user reviews. Completion becomes collaborative.
18 months+
Open the primitive layer externally. A2A readiness. Not a new project — an opening of what already works internally. One architectural decision unlocks all four phases.
06

The thought underneath all of this

"Software stops being a finished product delivered to passive users and becomes a capability substrate that different actors — humans, personal agents, external agents — each engage with on their own terms."

The piece of this that sits deepest: incomplete by design is not a bug, it's the new model. Software companies will intentionally ship primitives expecting users and their agents to complete the experience. The product is a substrate. The interface is a collaboration.

The UI as we know it was always a historical artifact of humans being the only actors. Once that assumption breaks, everything built on top of it must be rethought from primitives up.

The interface is not the product. The capability is the product. The interface is just one way of reaching it — and soon, not even the primary one.

Agentic AI A2A Protocols Product Strategy Software Architecture Future of UX
© Huy Do 2026