Third Inflection Is Happening Now

Imagine two builders in late 2025.

One is at a co-working space on Castro Street. He attends every AI Engineering meetup in SF. His bookshelf has the O’Reilly books on RAG, vector databases, agent orchestration. He is mass-subscribing to MCP servers, tuning embedding models, debugging chunking strategies. Notebook full. Slack pinging. Six browser tabs of documentation. He raged on twitter during the eval debate. He’s engineering complexity because that’s what serious AI work looks like. He refactored his memory system three times last month. He’ll probably refactor it again next week.

The other is going to Morocco from London, eating breakfast. Someone tweets about a bug in his code. He screenshots the tweet, sends it to WhatsApp, goes back to his eggs. By the time he’s finished eating, his agent has read the tweet, checked out the repo, fixed the bug, committed the change, and replied on Twitter that it’s done.

Same Opus model. Same API. Same documentation.

117,000 developers starred the second one. After four name pivots they call it OpenClaw. It’s the fastest-growing AI project of 2026.

It’s making people uncomfortable because it’s simple in a way that feels almost embarrassing. No vector database. No RAG pipeline. No orchestration framework. Memory is text files. The kind you’d edit in Notepad. Tool integration is CLI scripts, the kind you’d write in an afternoon and forget you wrote. The builder doesn’t use Anthropic’s MCP protocol. “I don’t use MCPs or any of that crap,” he said. He just wrote command line tools. One for Google APIs. One that reverse-engineered his Eight Sleep bed so the agent can cool his mattress before he gets in. One that hacked into his local Vienna food delivery app to tell him when dinner arrives.

His agent once watched his security cameras all night, taking screenshots every few minutes, convinced a stranger was sitting on his couch. It was a shadow. These things happen when you give an AI a cron job and tell it to pay attention.

Writers know this problem.

Ten people live through the same decade. Same city. Same heartbreaks. Same late nights wondering if this is all there is.

Nine write journals no one reads. One writes something that makes strangers cry.

The difference is never the life. It’s the tilt. The way you wear your hat. The way you sip your tea. Which details you hold. What you leave out.

Memoir isn’t about what happened to you. It’s about what you did with it.

The builder on Castro has more tools, more frameworks, more O’Reilly books. The builder in Morocco has a daemon that routes messages, a heartbeat that wakes the agent every thirty minutes, webhooks that listen, and files that persist. Every piece ordinary. The ordinary pieces, connected in a way that felt obvious only after someone did it.

Two things happened in 2025 that rearranged the business furniture of the technology world.

Claude Code shipped in March. An agent that wrote real code, in your terminal, on your codebase, while you got coffee. Agents stopped being slides in a deck. Then Opus 4.5 and GPT 5.2 dropped in November, within weeks of each other. Both crossed a threshold. Both became good enough for production work.

When every model is good enough, what’s left to compete on?

The harness. The wrapper. The nervous system around the brain.

Context is what a model sees. Harness is what it does with that. Context is memory. Harness is muscle memory. Armin Ronacher who created Flask, explains it well: “Harness is basically the thing that runs the loop and provides the tools and the prompts and all of that stuff.” Cursor, Claude Code, Windsurf, OpenClaw all run on the same underlying models. They feel like different species. You wouldn’t confuse them for a second.

If Claude Code feels like a colleague who happens to live in your terminal then OpenClaw feels like a weird friend who lives on your computer and texts you when your food delivery is late. Same brain. Completely different nervous systems. The harness is the product.

Your vendor tracker is a map of last year’s war.

Procurement watches Anthropic, OpenAI, Google, Meta. Compares benchmarks. Negotiates contracts. Worries about which model is ahead this quarter.

Meanwhile, the battle moved one layer up. Nobody updated the chart.

Developers are choosing ecosystems your vendor list doesn’t show. Cursor. Claude Code. Windsurf. OpenClaw. Lock-in is forming here, in the harness layer, and by the time a model hits your procurement tracker, it’s already commodity. You’re watching the engine. The race is being won by the chassis.

Cloud commoditized servers. Value moved to orchestration. AWS won.

Mobile commoditized distribution. Value moved to apps. Apple won.

AI is commoditizing models. Value is moving to harnesses. I don’t know who wins this one. But I know where to look.

The builder on Castro is still collecting RAG tutorials. More components. More dependencies. More failure modes. More things to debug at 2am. His architecture requires a whiteboard to explain, and even then, you’d squint.

The builder in Morocco is tinkering. Trying things. Seeing what works. Throwing away what doesn’t. His architecture fits in his head. He could explain it to his nephew.

KISS. Keep it simple, stupid. Engineers have known this for decades. Complexity is easy. Anyone can add a component. Simplicity requires taste. Knowing what to leave out. That’s harder, and it’s why most people don’t do it.

Somewhere in your org, a developer has a side project running on a cron job. It’s not on any roadmap. It probably works better than most other things.

117,000 stars for markdown files and a cron job.

Sometimes the answer was always obvious.