# What Agents Want From Your Product

Agents don't want your UI. They look for an API, an MCP server, a CLI, or a skill, and they pick the products that give them one.

## The shift from clicking to delegating

Your next wave of users won't click through your product. They'll delegate the task to an agent and expect it to get done. That agent won't sit through your onboarding flow, follow your navigation, or read your marketing copy. It will look for a way in it can use.

## The 4 things agents want, and the one they don't

Agents look for four surfaces: an **API**, an **[MCP server](/glossary/mcp-server)**, a **CLI**, or a **skill**. The one they want least is your **UI**. If none of the four exist, the agent either drives a browser through screens built for people or picks a competitor that gave it something to call.

This isn't a ranking of which surface is "best". Different agent types reach for different things. A coding agent working in a terminal will happily use a CLI. A hosted assistant like Claude or ChatGPT reaches your product through an MCP server, listed as a connector, app, or plugin. An autonomous, long-running agent may need the API, the CLI, and the MCP server, plus a skill that packages the workflow so it doesn't have to rediscover it every time. Claude and ChatGPT now bundle MCP servers and skills into plugins, so a product can ship both as one install. That is the basic shape of the [Agent Access Stack](/guides/agent-access-stack-apis-cli-mcp-skills).

## Why the UI is the weakest option

Most products still lead with their UI, because that's what humans need. But a UI assumes a visual session, a human clicking buttons in sequence, and a person who can read error messages and course-correct. None of that comes naturally to an agent. If your only door is the UI, agents fall back to driving a browser. Claude in Chrome can now click, type and fill forms with a person's login, and Anthropic pitches it for the "internal dashboards, legacy systems, and vendor portals" that don't connect any other way. You don't want your product in that group.

## What good design for agents looks like

A few principles show up once you start building for agents instead of bolting AI features onto what you have. The model controls the flow, not you. That's the biggest mental shift from designing for humans. A tool's description isn't documentation. It's an instruction the model reads to decide whether to call the tool, and vague ones get skipped even when the tool is exactly right for the job. Good instructions say when to use a tool and when not to: negative cases matter as much as positive ones. Many existing APIs need real redesign for agent and conversational use, not just a wrapper. And agents do better with one bulk call than with the same call twenty times in a row. Start that redesign with the [AI-ready API checklist](/resources/ai-ready-api-checklist).

*(Several of these principles come from Elliot Garreffa and Hanh Nguyen's [lessons from building 100+ ChatGPT apps](https://elliotgarreffa.beehiiv.com/p/18-lessons-from-building-100-chatgpt-apps-with-fractal-s-co-founder).)*

## What to do next

Audit your own product the way an agent would encounter it. Does it have an API an agent could call blind? Does it have an MCP server, or a CLI, or a skill wrapping any of that? If the honest answer is "just a UI", that's the gap to close first. Start with the API: everything else in the stack builds on it. If you have hundreds of APIs, read [how to determine which APIs to make agent-ready](/guides/how-to-determine-which-apis-to-make-agent-ready) before you build any wrappers.

---

Want a full benchmark of your APIs and MCP with prioritized fixes? [Get in touch](/contact).

### FAQ

**My product only has a UI right now. Where do I start?** Start with your API, even a minimal one. MCP servers and CLIs both wrap an API. Without one underneath, there's nothing for them to expose. A UI-only product leaves agents clicking through screens built for people, no matter how good those screens are.

**Do I need all four surfaces: API, MCP, CLI, and skills?** No. Different agent types use different interfaces. Chatbots, coding agents, and autonomous agents don't all reach for the same thing. Start with a solid API, then add MCP or a CLI based on who's going to consume it, and layer a skill on top once there's a workflow worth packaging.

**We have hundreds of APIs. How do we know which ones to make agent-ready first?** Start with evidence, not intuition. Run a real agent against your top candidates and watch where it guesses, retries, or gives up. Look first at the APIs your customers already automate around, and don't skip internal or niche ones: models haven't seen them in training, so they fail more often.

Source: Level 250 — https://www.level250.com/guides/what-agents-want-from-your-product