Applied AI: Three Learnings from Shipping — My Talk at the WorkOS Applied AI Showcase

The WorkOS Applied AI team did a live showcase today. The first hour was demos — our autonomous agent harness CLI, the GTM intelligence platform, and the Horizon agentic coding platform. I closed it with my talk: Applied AI: three learnings from shipping.

My segment runs from 1:02:45 to about 1:17:23. The embed below jumps straight to it.

Play the deck

Three things, one motion

The frame for the talk: three things I've learned shipping AI internally at WorkOS over the last eighteen months. They're not independent — they compound into one motion.

1 · Interface beats stack

The same idea, two interfaces, eighteen months apart.

In 2024 I built Bartleby — a standalone web app that did most of what Blog Bot does today. Almost nobody used it. Required a separate login. Required a separate URL. Required a separate context-switch from whatever the user was already doing. The verdict was a moment of silence, the kind of silence Bartleby the scrivener would have appreciated: "I prefer not to."

In 2026 I shipped Blog Bot — a Slack bot you @-mention in the channel you already live in. Tag it in a thread, it writes the draft, you ship. Required zero new behavior.

Same model class. Same prompts. Same outputs. Different door. The 2026 version got adoption immediately because it removed the only thing that actually mattered: friction at the point of use.

Under the hood, the stack is anything but simple. Every @-mention spins up a Cloudflare Workflow with 20+ durable steps — route the conversation, pull Granola notes, fetch YouTube transcripts, match a format reference. The asymmetry is the leverage: one @-mention in, a fully drafted 1,635-word post out, five minutes later.

The lesson: meet users where they already live. The interface is the product.

2 · Complete the loop

Capture → plan → execute → ship.

I work against one Linear ticket per project. Subtasks for bugs and features. The agent /loops against the main ticket — picks up the next subtask, implements it, opens the PR, deploys, marks it done, picks up the next one.

Three integrations, three minutes from idea to in-progress: Linear → GitHub → Slack. The kit is internal, the heuristic is the point.

Speed requires safety. The loop runs with guardrails — every change goes through a verified, deployed sandbox before it touches main. A bug becomes a deployed change without ever touching production directly. I review. The bot does the typing. Two human touches per fix.

One opinion that came up in the room: give your agent the tools directly, not via MCP. MCP is great for cross-tool portability, but for your own internal harness, direct tool calls are faster, easier to debug, and easier to constrain. Save MCP for the boundary where another agent needs to reach in.

3 · The imagination gap

Model capability outruns our product imagination. Always. By a wide margin.

The exercise: re-ask the question every quarter. "What would I build now that wasn't possible last quarter?" The answer is never zero. This month's answer for me was a long-running-thread pattern I'm watching emerge.

The pattern: let humans talk. Flush on request.

1. Listen — keep the buffer. Every message in the thread streams into D1. No edits, no opinions, no interruptions.
2. Wait — let humans collaborate. People @-mention each other, debate, change their minds. The bot doesn't care — yet.
3. Flush — act on request. "Hey bot, apply the edits we just discussed." It reads the buffer, makes the change, ships.

It's the inverse of every chatbot pattern of the last three years. The bot is a journal that activates on cue, not an interrupter. Once you see it, you see it everywhere.

Why this transfers

If you're trying to roll AI tooling out at your own org, these three are the ones I'd give you in priority order. The interface decision determines whether anyone uses what you build. The loop decision determines whether what you build compounds. The imagination decision determines whether you're still building the right thing in six months.

If you want to talk about how this might work at your org, book a workshop or drop me a line. The Applied AI team runs workshops based on exactly this material.