Happy Friday.
Last Friday I published This is the New Type of Design Work, my attempt to describe what’s happening with agentic design.
And then Tuesday’s podcast became our most discussed episode as Diego Zaks, head of design at Ramp, broke down how their design function works as a 1,500-person company that has a 99% AI adoption rate.
Today, my OpenClaw for Designers video just went live. It’s the most honest stack walkthrough I’ve ever shared, and it cost me over $13k to build.
For now, let’s talk about how to identify an agentic setup.
– Tommy (@designertom)
JOIN OUR DISCORD
Today, our private Discord opens for one week.
It’s a community of designers learning to build and sharing lessons. But we’re also a slightly unhinged group of people who have become friends.
We kick purge inactive members every quarter. We do not tolerate spam.
And on Monday, we kickoff a Mac Mini agentic design workflow challenge. Show off an AI-powered design workflow or custom tool. The community picks a winner. The winner takes the Mac Mini.
What actually passes the test for me
A chatbot is a window. An agent is a loop.
Answer these four questions to determine if your chatbot passes the agent test:
What can it do repeatedly? I have skills the agents run on their own. Audit a design against my taste rules. Pull research into a brief. Run my Friday cleanup. Process tagged content into the chronicle. None of those is impressive in isolation. They just happen the same way every time, without me rebuilding the prompt.
What does it remember? A vault of operating files. SOUL.md tells the agent who I am and how I think. USER.md tells it what I am working on. MEMORY.md is the running log of decisions and corrections. And my knowledgebase powered by Obsidian. Without those, the agent forgets the same correction I made last Tuesday and I have to make it again.
What can it see right now? Whatever I point it at. The Slack thread. The Obsidian note. The transcript I just dropped in. The Figma file via MCP. Most bad AI output happens because the model is guessing inside a tiny window. When you expand the window, the outputs improve drastically.
What can it reach? Slack, Obsidian, GitHub, Linear, my calendar, Beehiiv, a browser, a terminal. This is also the layer where I slow down. Connections have keys, and those keys touch private data, so I started narrow and only expanded after the loops earned it.
What it actually costs
I spent over $13,000 in Anthropic charges over the last 90 days, across 68 invoices. Roughly $4,400 a month in tokens this year.
The bet is real, and so is the buy-in.
Tuesday I sat down with Diego Zaks for State of Play. He runs design at Ramp, and his line on the show was: "You can’t be a designer here and not be using AI. It’s just not a thing."
Karri Saarinen at Linear posted his own receipt: Agent-delegated work climbed from 10% in February to 24% in April.
We’re starting to hear more evidenced reports back from the first AI installers at scale. And while it’s not clear if it’s “working”, it’s adoption curve continues to climb.
What gets you in trouble
If your agent is making you ship faster but review less, you built a slop cannon.
Hannah Ahn at Superpower said it on a recent call: "let’s reel it in." Her team has more reviews now than before, not fewer. That is the right shape: speed up parts of the producing, keep the gating manual.
If you’re feeling beat down by AI, me too. And while it’s not slowing down, I am.
I’m taking one day a week off AI-powered work, and I’m still deep in post-production of our Colin & Samir documentary about the rise of “home movie-makers”, and this soulful story is giving me life.
The Monday move
Pick one workflow you already own. Map every step. Mark what an agent would need to do, remember, see, and reach. Give it one job. Keep yourself in the approval loop until the loop earns its trust. And even then…
The OpenClaw for Designers walkthrough is up now if you want to know what an agent harness is, what’s in mine, and how to build one.
Tell me what you think of it in the comments.
Tommy
