Last week, in a cramped office on Market Street in San Francisco, 405 people squeezed into a room to watch a series of demos. No slides, no corporate pitches—just five minutes of raw, live coding.
WorkOS founder Michael Grinich stepped up first. He showed off a fully functional, keyboard-driven LinkedIn messaging client. It had a sync engine, offline capabilities, and over 900 tests. It was roughly 15,000 lines of TypeScript.
He wrote zero of it. Claude did. He sat by a pool on Memorial Day weekend and let the AI build the entire thing. He didn’t read a single line of the output.
When a developer brags about shipping 15,000 lines of code without reading a single one, the paradigm hasn’t just shifted—it’s already left you behind.
The most terrifying part of the night wasn’t the tech. It was the reaction. The room didn’t gasp. They just nodded, bored, as if this was the expected baseline. That collective shrug is the loudest alarm bell in the software industry right now.
You’ve probably felt it. You’re using AI agents to write code 10x faster. You’re crushing tickets. But your team isn’t shipping 10x faster. The bottleneck didn’t disappear; it just moved.
It moved to code review. You cannot ask a human to review 15,000 lines of AI-generated code. Line-by-line review is officially dead. Traditional workflows are breaking down because we are trying to force human collaboration models onto machine-generated output.
We are still trying to review the output, while the real value is locked in the conversation that created it.
Think about how you use Cursor or Claude Code right now. You prompt it, you debug with it, you argue with it, and then you close the terminal. You save the code to GitHub, and you throw the chat history in the trash.
That is a catastrophic mistake.
The most provocative realization from this Demo Night is that the code itself is no longer your company’s primary asset. The AI’s session history, decision logs, and context are. The code is just a compiled byproduct of the actual engineering—the reasoning, the trial and error, the context that happened in the chat window.
If your AI agent’s memory resets every time you close the terminal, you’re burning your company’s most valuable IP.
Startups are already figuring this out. A company called Entire is tying agent sessions directly to git commits. Every commit becomes a ‘checkpoint’ that saves the agent’s intent, the tools it used, and the token cost. When a new engineer joins, they don’t just read the code—they search the past agent sessions to understand why the code exists.
Brian Douglas showed off a tool where he forced Claude to play Pokémon. But the real magic was his ‘tapes’ feature: recording the high-fidelity trajectory of the agent’s failures and successes, and feeding that memory back into the next session. He’s doing this for internal engineering, feeding past sessions to agents so they can automatically set up test environments for new hires.
Meanwhile, tools like Modem’s SideShow are translating massive, unreadable AI pull requests into plain-English web pages with badges and visualizations. The code diff is no longer a communication medium between developers; it’s a communication medium between humans and agents.
If you work in software, your current tools for code review and team communication are already obsolete.
The teams that thrive in the next two years won’t be the ones who write the best prompts. They will be the ones who figure out how to make AI agents true collaborators—agents with persistent memories, shareable contexts, and decision logs that outlast a single terminal session.
The code is just the exhaust. If you aren’t capturing the engine’s telemetry, you’re driving blind. Stop obsessing over the lines of code. Start obsessing over the context that generated them.
FAQ
Q: Isn't AI code still just code that needs human review?
A: No. 15,000 lines of unreadable AI code is a liability, not an asset. Human line-by-line review at that scale is a myth. You have to move alignment upstream to the planning phase.
Q: What do we do with our current Pull Request workflows?
A: Stop forcing humans to read massive diffs. Use tools that let agents explain their changes in plain English, and tie the AI's chat session directly to the commit so the context is preserved.
Q: Are you saying we shouldn't even read the code our companies ship?
A: I'm saying the code is a compiled artifact. The actual engineering happened in the prompt and the debugging session. If you don't save that context as a first-class asset, you lose your company's IP.