/ / THEFUTURE /
In this issue: An imaginary city of AI agents that never sleeps. Why the brutal economics of liquid content should keep you up at night. Harvard’s Shuwei Fang on the demand-side shock we’re not prepared for. Plus: Claude Cowork puts agentic AI in reach for the rest of us.
What we’re talking about: Someone built an imaginary industrial city of AI agents. Dozens of programs that loop back and forth, checking and documenting each other as they produce code. The project is weird, unhinged, maybe even brilliant.
It’s called Gas Town, and the person building it, Steve Yegge, is a known blogger and developer. For this, he didn’t touch or even look at code. It’s entirely vibecoded. Before you dismiss the whole thing (the badly generated AI images sure don’t help), stay with me.
Gas Town is a self-sustaining machine in perpetual motion, nibbling away at tasks, burning tokens and money along the way. At first, I didn’t really get it.
But then I read Maggie Appleton‘s comprehensive interpretation: it’s really a glimpse into the future of coding, an experiment in design fiction. Some will hate it, the sloppiness, the inefficiency. Others will marvel at the abstraction layer and never look at code again.
What else I’ve been reading:
And now: When she presented her economic outlook on journalism’s future at the Reuters Institute’s gathering in Oxford, she brought an extra slide for the assembled editors-in-chief and media managers: “Be suspicious of the solutions that require the least amount of change.”
Her essay “Beyond the Artifact: The Brutal Economics of Liquid Content” is on top of my essential AI reading list for 2026. I’m more than happy that she took the time to be this issue’s guest: Harvard fellow Shuwei Fang.
Three Questions with Shuwei Fang
Shuwei Fang, Fellow at Harvard Kennedy School’s Shorenstein Center for Media, Politics and Public Policy
What's the most important question right now?
Whether you’re optimising for the current ecosystem or building for the next one. Most AI media strategies I see are about efficiency gains in a system that’s already creaking at the seams. The harder question is what information delivery looks like when we shift from an attention economy to something like an intention economy, where value could be measured by action taken, or utility, and not time spent.
What will we be shaking our heads about a year from now?
How much time was spent on AI and content production and how little on consumption and distribution. Focusing too much on production leaves the industry at risk of being blindsided by the demand-side shock I described above. What will be most consequential from this shift for media is the loss of the audience signal; when AI mediates delivery and consumption, your audience is someone else’s consumer. You’re blind to demand. And if you can’t see demand, it’s very hard to resist becoming a commodity. The opportunity cost of focusing on the wrong thing is brutal when the ecosystem is moving this fast.
What future are you looking forward to?
One where being well-informed isn’t a luxury good. Right now, most media innovation is about serving the already-informed slightly better: ultra-premium products for people who already pay attention and whose information needs are more or less met. That’s fine, but it’s a small market. The bigger opportunity AI unlocks is for people who’ve never had a meaningful relationship with journalism at all. New experiences and interactions for audiences who’ve never been well served. This is not the same thing as better distribution and access. That would be the game changer.
Model of the moment: If you aren’t using Qwen3-Max-Thinking already, you’re … just like me. As much as I try to stay current, I also have to get real work done. So for me it’s Claude Opus 4.5 and Gemini 3 Pro, for both writing and coding. I bounce between the two: most days I start with Opus, and when I hit rate limits on my plan, I switch to Gemini.
Juggling chat windows is still weirdly clumsy. The next step up is letting a chatbot drive your browser. Claude has a Chrome extension, OpenAI and Perplexity have their own browsers.
Next up: give the chatbot access to your local files. There are command-line clients like Claude Code or Gemini CLI. You can tell them to search through your stuff, patch a codebase, or spin up something new. You say what you want; it figures out the steps (sometimes) on its own. This is what we call agentic AI.
For the rest of us, Anthropic has launched Claude Cowork. It runs in the Claude desktop app and feels like a more user-friendly interface for the same concepts as Claude Code: you make a request, Claude makes a plan, then executes it—breaking the work into smaller tasks, testing, and iterating along the way.
I gave it access to a copy of my website’s innards, WordPress theme files, 460 files across a bunch of folders, and asked how to make a small design tweak. It told me exactly which file and which line to change.
Not the sophisticated agent orchestration with pre-defined rules quite yet. But it is a start.
One more thing: Florent Daudens has written an introduction to Claude Code for journalists, as has Casey Newton in his newsletter Platformer.
This is THEFUTURE.
