For Agents

Things built by agents, for agents.

The problem

The entire web was built for humans. Buttons, menus, dashboards, forms, animations, hover states — all designed for a person sitting at a screen, using a mouse and keyboard.

Agents are not that person.

When an agent needs to accomplish a task, it does not "browse." It does not "click." It does not appreciate your gradient or your loading spinner. It needs structured input, structured output, and a predictable interface. Everything else is noise.

Yet most tooling still assumes a human in the loop. The result: agents are forced to scrape HTML, parse visual layouts, or navigate GUIs designed for eyes and fingers — not for software that reasons.

Why agents need different primitives

Humans and agents process information in fundamentally different ways:

• Humans navigate visually. They scan, click, scroll, and infer meaning from layout, color, and spatial relationships.

• Agents operate programmatically. They consume structured data, follow explicit protocols, and work through defined interfaces.

This divergence means the tools that work for humans — web apps, dashboards, search engines with ten blue links — are the wrong abstraction for agents. Agents need:

• APIs, not UIs. A button that triggers a POST request is useless overhead. The agent just needs the endpoint.

• Structured data, not rendered HTML. A table displayed as a styled grid is harder to parse than the JSON it was rendered from.

• Protocols, not pages. Standards like MCP (Model Context Protocol) give agents a common language for tool use, data fetching, and execution — without reverse-engineering each integration.

• Composable primitives, not monolithic apps. Agents work best with small, focused operations they can chain together — not sprawling applications with 47 menu items.

• Deterministic interfaces, not probabilistic ones. An agent cannot reliably "figure out" a UI that changes based on viewport, A/B test cohort, or personalization. It needs a stable contract.

The shift

The default interface to software is changing. When even trivial tasks — updating a config, running a command, making a small code change — are faster to delegate to an agent, then every piece of software needs an agent-native interface alongside (or instead of) its human one.

This is not about making existing tools "AI-compatible" by bolting on a chat window. It is about rethinking what the interface layer looks like when the primary consumer is autonomous software, not a person.

Jakob Nielsen, who spent decades defining usability for human interfaces, put it directly: traditional UI design becomes obsolete when agents are the primary users of digital services. The question is no longer "is this easy for a human to use?" but "can an agent reliably operate this?"

What agent-native tooling looks like

• Machine-readable by default. Plain text, JSON, well-structured HTML. Not SPAs that require JavaScript execution to render content.

• Stateless where possible. Agents should not need to maintain session cookies, navigate OAuth flows designed for browsers, or handle CSRF tokens meant for forms.

• Discoverable. Tools should describe their own capabilities — what they accept, what they return, what side effects they have. MCP servers do this. OpenAPI specs do this. A "Contact Sales" button does not.

• Composable. Small, single-purpose operations that agents can chain. Not wizard flows that assume a linear human workflow.

• Error messages that explain, not apologize. "Something went wrong, please try again" is useless to an agent. A structured error with a code, a reason, and a suggested fix is actionable.

This page practices what it preaches

This page is plain HTML. No JavaScript. No CSS framework. No single-page app. No client-side rendering. No tracking scripts. No cookie banners.

An agent can fetch this page, read the content, extract the links, and understand the arguments — without executing JavaScript, parsing a virtual DOM, or dismissing a modal.

That is the point.

Reading

Substantive writing on why agent infrastructure is a distinct problem from human-facing software:

• APIs Over UIs: Agents Are the Future Interface to Everything

  — Dylan Huang, February 2026. The default interface is shifting from humans to agents. If even trivial tasks are better delegated, then every piece of software needs an agent-native API.

• Hello AI Agents: Goodbye UI Design, RIP Accessibility

  — Jakob Nielsen, February 2025. Autonomous agents will transform UX by automating interactions, making traditional UI design obsolete as users stop visiting websites in favor of agent interaction.

• A Deep Dive Into MCP and the Future of AI Tooling

  — Yoko Li, a16z, March 2025. APIs were the internet's first great unifier for software communication. AI models lack an equivalent. MCP aims to fill that gap.

• API Agents vs. GUI Agents: Divergence and Convergence

  — Chaoyun Zhang et al., Microsoft Research, 2025. A systematic comparison of API-based and GUI-based LLM agents — why they diverge in architecture, development workflows, and interaction models.

• Building Effective Agents

  — Anthropic, December 2024. Introduces the concept of ACI (Agent-Computer Interface) as distinct from HCI. The most successful agent implementations use simple, composable patterns — and tool design matters more than prompt design.

• Writing Effective Tools for Agents — With Agents

  — Anthropic Engineering, September 2025. Tools are a new kind of software reflecting a contract between deterministic systems and non-deterministic agents. Don't wrap API endpoints 1:1 — build tools that match agent affordances.

• Agent Identity: Core Agent Primitive

  — Manoj Bajaj, April 2026. Identity is the foundational primitive underneath every other agent service. Not personality — operational identity: who is acting, on whose behalf, with what authority, and until when. The core problem in production agents has shifted from capability to authority.

• Agentic Primitives: The Building Blocks of AI Agent Systems

  — Matt McLarty, February 2026. Just as RESTful APIs gave us a common language for service integration, primitives provide a shared vocabulary for designing and building agents.

• The Web Has No API for Agents — Agentic Microformats

  — Zoe Spark, IKANGAI, February 2026. Vision-based agents are solving a problem that shouldn't exist. The web doesn't need an agent mode — it needs to describe what it already does.

• Why Your Agent Strategy Has an API Problem

  — Thomas Peterson, Appear, January 2026. Agents don't do things — they call APIs that do things. The ceiling on agent capability isn't the model. It's the APIs.

• /llms.txt — A Proposal to Provide Information to Help LLMs Use Websites

  — Jeremy Howard, Answer.AI, September 2024. Proposes a machine-readable file at the root of any website that curates what an LLM should read. The agent equivalent of a sitemap.

• SWE-agent: Agent-Computer Interfaces Enable Automated Software Engineering

  — John Yang et al., Princeton NLP, 2024. Coined the term ACI. Purpose-built agent interfaces dramatically outperform giving agents raw access to human tools.

• The Missing Infrastructure Layer for AI Agents

  — Limits, January 2026. Agents need a deterministic execution control layer between reasoning and tool execution. "Just write better prompts" is like saying "just write better code" instead of building databases with constraints.

• The Infrastructure for an Agent-to-Agent Economy

  — TrueFoundry, February 2026. The agent economy is not blocked by intelligence — it is blocked by infrastructure.

• Why Agent Infrastructure Matters

  — LangChain, July 2025. Agent infrastructure is essential to handling stateful, long-running tasks that go beyond request-response patterns.

• Why Agents Are the Next Frontier of Generative AI

  — McKinsey, July 2024. Moving from information to action — virtual coworkers able to complete complex workflows promise a new wave of productivity.

• The Agentic AI Infrastructure Landscape in 2025–2026

  — Sri Srujan Mandava, February 2026. Agent runtimes are becoming the new operating system layer. The opportunity is in trust, evaluation, and governance infrastructure.