I started treating AI agents as customers, and it changed how I build. An agent

By Rodms10 · October 20, 2026 · Curated by George's Blog

I started treating AI agents as customers, and it changed how I build.

An agent starts knowing nothing about your new product. It learns your tool from scratch in every session, so a confusing first experience costs time and money every single time.

Velloo's prototype gave agents 77 tools and about 27,300 tokens to read before their first step. That wasn't viable. So I looked for what a good customer experience would look like for an agent. Today it starts with 4 tools and about 2,650 tokens, and picks up the rest as it works.

That's one of the techniques in my new post on building an MCP server for agents that have never seen your tool:

→ Load less. A few stable tools in front of every operation, guides fetched on demand, and a token budget that fails the build.

→ Write for what the client shows. Claude Code cut our instructions short and hid our tool descriptions behind a search, so we reordered the brief and load the main tools upfront.

→ Use a language the agent already knows. Agents compose screens in JSX, not a JSON format we invented.

→ Accept a reasonable guess. An argument with only one possible reading is accepted, and the result says what was read. Everything else gets one plain sentence and the schema.

→ Return only what helps next. No echoed screens, unchanged warnings sent once, and comparisons that name the element to fix.

→ Measure first calls, not just final screens, across models and clients.

Then I asked two AI agents to review my product. Not a benchmark score: a customer review, in their own words, at the end of a test design session.

Opus 5.5 in Claude Code: "…one call gave me the real page built from the app's own components…"

Sol 6.1 in Codex: "Velloo made recreating the full page surprisingly quick…"

They also told me what got in their way. Some of those complaints turn out to be real bugs that then get fixed based on detailed reports with reproduction steps.

If you're building an MCP server, try asking your agents what it was like to use it. You'd like to hear what they say.

https://lnkd.in/eWKta78Z

View the original post on LinkedIn