Are Agents Really Eliminating UI?

While predictions suggest that AI agents will replace software UIs, in reality, UIs are not disappearing but rather diverging in usage. The form involves adding a headless layer where agents interact directly with products on top of the screens used by humans.

The first entry point for tech products may shift from homepages to terminals or AI chat windows. This is because if users receive product recommendations from Claude or ChatGPT, agents can handle sign-up and configuration via MCP. This requires onboarding that agents can easily find and authenticate, a setup process completed in a single call, parseable documentation, and short, scope-limited tokens.

However, even after agents complete setup, humans must open the product to verify completed tasks and next steps. A UI is still required for the process of reviewing or approving results generated by agents.

Products must design for the following two entry points together.

• For agents: End-to-end sign-up flows, connector lists, llms.txt and standard Markdown documentation, secure authentication methods
• For humans: Screens that show how far the agent has progressed and guide users on what to do next

PostHog's npx -y @posthog/wizard@latest sets up an account, identifies instrumentation targets in the codebase, and logs progress to the terminal. Afterwards, some users continue working via MCP, but many return to the web app to verify and explore data.

Even if a Slack app converts message prompts into draft PRs, users check the diff on GitHub before merging. AgentMail also allows agents to create email accounts and keys without a console, but a human UI remains for the step of delivering the 6-digit code sent to the signer-up.

The computer-use approach where agents directly click UIs is possible but slow and closer to a human-centric experience. Instead, AX (Agent Experience), which explains what a product can do and allows direct manipulation, is important, with MCP currently being the representative standard.

To check if a product is agent-friendly, first run it in a headless environment and identify where human intervention is required. The key lies not in eliminating UI, but in designing a hybrid experience used by both agents and humans.