The Three Layers of Modern Software: UI, API/MCP, and Agents

Software has historically been defined by its user interface: log in, navigate to a screen, fill out a form, and read a report. That model is not disappearing. It is becoming one layer of a larger platform.

The next generation of software is being built across three connected layers: a UI for people, an API/MCP layer for systems and models, and an agent layer that can carry intent through to useful work. This is a thesis about architecture, not a claim that every product has already arrived. The evidence, however, increasingly points in the same direction.

The market is moving beyond UI-only products

Forrester’s 2026 enterprise software outlook is unusually direct. For decades, enterprise software was designed primarily to assist people. Forrester argues that applications must now accommodate a digital workforce of AI agents, shifting from a user-centric design philosophy toward one centered on workers and processes. Tech leaders will need to decide how far to digitize processes and orchestrate workflows that do not depend on someone clicking through every step.

That shift does not make APIs optional; it makes them more important. In a separate architecture note, Forrester argues that AI struggles at the lowest levels of abstraction and performs better when it can work with clear, business-level services. Fragmented, UI-bound products therefore become liabilities, while products with durable APIs become building blocks that agents can assemble. Forrester also identifies a growing need for real-time context. Standards such as the Model Context Protocol point toward a layer that lets agents discover, understand, and securely invoke enterprise capabilities.

The agent layer is no longer just a lab experiment. Gartner predicts that 40% of enterprise applications will feature task-specific AI agents by the end of 2026, up from less than 5% in 2025. It also warns against “agentwashing”: marketing assistants that still depend on constant human input as agents. Gartner’s longer-term forecast is even closer to this three-layer model. By 2028, it expects agent ecosystems that let users achieve goals without interacting with every application individually—an outcome that requires applications to expose more than a screen.

MCP is one piece of the connective tissue behind that forecast. Anthropic introduced it in November 2024 as an open standard for connecting AI applications to the systems where data and tools live. A year later, the Linux Foundation announced the Agentic AI Foundation, with Anthropic donating MCP as a founding project alongside OpenAI and Block. The Foundation described MCP as a universal protocol for connecting models to tools, data, and applications, with more than 10,000 published servers and adoption across Claude, ChatGPT, Gemini, Microsoft Copilot, Cursor, and VS Code. Bloomberg’s supporting quote captures the architectural significance: MCP is a foundational building block for APIs in the era of agentic AI.

Forrester’s Predictions 2026 report then closes the loop for software vendors: 30% of enterprise application vendors will launch their own MCP servers so external agents can collaborate with their platforms. In that model, the MCP server works with the platform’s APIs, and agents can access and act only on authorized data, just as a human user would. This does not replace the UI. It adds second and third interfaces to the same governed product.

Software used to end at the screen

The first generation of cloud software replaced local installations with browser-based applications. The next generation made those applications easier to connect through APIs, webhooks, identity standards, and data exports. Both were important advances, but most products still treated the UI as the product and everything else as an integration project.

The next shift is about making a product useful beyond a person clicking through a workflow. A system can now be accessed by a campus integration, a data pipeline, an AI client, or an agent coordinating several steps. That does not make the UI less valuable. It clarifies what the UI is responsible for: trust, context, exceptions, collaboration, and approval.

The best software platforms will not force a choice between human usability and machine usability. They will provide a coherent foundation for both.

The three layers

API and MCP are complementary

It is tempting to describe MCP as a replacement for APIs. That is the wrong abstraction, and the market evidence does not support it. REST APIs remain the durable contract for resources, authentication, writes, integrations, and system-to-system operations. Webhooks let connected systems respond when authoritative data changes. Forrester’s own prediction assumes that an MCP server works with a platform’s APIs rather than around them.

MCP adds a model-oriented interface. It gives an AI client a discoverable way to access selected context and tools while the product retains control of authorization and the underlying data. The distinction matters because models do not interact with products the same way integration services do. They need useful context, explicit tools, and clear boundaries around what they can do.

The API/MCP layer is the connective tissue between the product and the rest of the software ecosystem. It makes the platform available without requiring every new client to reproduce the UI.

One platform, three ways to work

The three layers are valuable only when they share the same foundation. A UI that shows one version of the record, an API that exposes another, and an agent that reasons over a third creates more operational risk, not less. That is why Forrester tells buyers to question vendors about MCP and governance, and why Gartner distinguishes agents from assistants. The architecture works only when identity, permissions, and auditability travel with the work.

How HealthTasks is building toward all three

HealthTasks is a clinical education platform, so its architecture must support consequential work across schools, programs, and clinical partners. Today, the platform already operates across all three layers.

UI layer. Students, faculty, coordinators, and leaders use HealthTasks to manage clinical education, review evidence, coordinate placements, and understand program performance. The interface remains where teams inspect context and make decisions.

API/MCP layer. The HealthTasks REST API and webhooks connect clinical education data with institutional systems, reporting tools, and clinical partners. Our institution-scoped MCP endpoint makes governed HealthTasks context and tools available to AI clients.

Agent layer. HealthTasks Agents reason over live institutional data to answer questions, generate analytics, surface accreditation evidence, and prepare operational summaries. Agent Actions extend that capability into placement work, where a coordinator can review and approve proposed changes before anything is saved.

These are not three disconnected products. They are three access patterns over one connected clinical education system. That is the platform we are building and the direction in which we will continue to invest.

What platform buyers should ask next

A feature checklist is no longer enough to evaluate a software platform’s long-term architecture. Buyers should ask how the product behaves when a person is not the only consumer—the same question Forrester recommends asking enterprise application vendors as MCP and agentic workflows move from experiment to procurement.

• Can people still see the underlying record and review consequential decisions?
• Can our authorized systems read and update the data through documented interfaces?
• Can AI clients access useful context and tools without bypassing institutional controls?
• Can agents prepare or carry out real work instead of only returning a paragraph?
• Do the UI, integrations, and agents share permissions, source data, and an audit trail?

The future is not less software. It is more accessible software.

The future is not a world where everyone abandons the application for a chatbot. Nor is it a world where APIs and agents make product design irrelevant. The strongest platforms will serve different kinds of users through different surfaces while preserving one trusted system underneath.

People will continue to need a UI. Institutions and software will need APIs. AI clients will need MCP-style context and tools. Teams will increasingly expect agents to move from an answer to a reviewed action.

HealthTasks is already built across these three layers. We will keep investing in the foundation that makes them useful together: trustworthy clinical data, governed access, model flexibility, visible reasoning, and human control over consequential work.

References