The Drip
Everyone shipped an "AI coworker" recently. OpenAI released ChatGPT Work, Microsoft is standardizing Copilot on GPT 5.6, Meta launched Spark for agent orchestration, and Gemini rolled out new features of its own. It feels like an arms race in the news cycle, and it probably is. Every lab is competing to figure out what people using AI actually want, now that "go chat with it" has turned into "do real work with it." Our take: the products are converging fast. The names are not. As Kellan put it on the show: the nouns are getting out of control.
Inside The Bottle
Here's the thing nobody launching an AI coworker this week wants to say out loud: cowork, code, Codex, and ChatGPT Work are effectively the same thing. Harnesses for multi-step tasks. That's it.
The product packaging can get confusing. These companies are serving two audiences at once: devs who live in Claude Code and Codex, and everyone else who just wants a chat box. The result was that "agentic capabilities for business people" had no obvious home. So, we enter the Coworker craze. It's hidden behind tabs, pop-ups, and product names that describe org charts more than they describe what you can do.
Kellan's hot take: get rid of the tabs. One surface, a chat bar for quick questions, and let the model figure out whether it's looking at a code problem or a personal one. The models are smart enough now to deploy agents when they're needed. The tabs exist to make two user bases happy, and they're confusing both.
Until that happens, here's the practical way we split our own work:
| ■ | Runtime → Claude Code. If the thing you're making will run on its own (an app, an automation, even a script), use the harness built around writing code. |
| ■ | Shared context → a harness pointed at your knowledge. Work that leans on what your business knows: documents, plans, materials. |
| ■ | Quick answers → chat. A recipe, a golf trip, a fast question. No harness required. Chat isn't going anywhere. |
The bigger point: making the leap from a chat-based AI experience into an agentic one is the leap that matters right now. It's really beneficial when you start figuring it out in your own way, whatever the tabs are called that week.
Lab Notes
| ■ | Justin's note: I audited my own setup this weekend and found real sprawl. The fix I'm working toward: clear, simple domain-based entry points. Simple folder structure, as few external tools as possible. What I actually want is to open Claude and say "it's time to work on marketing" and not think about what's living where. |
| ■ | Kellan's note: Cowork and code are basically the same harness, but cowork typically gets more tokens, because that's the audience the labs are courting. Same capabilities, maybe 2x the tokens. That won't last forever. Use it while it does. |
What Stopped Our Scroll
|
