The bottleneck was never the code
Everyone's talking about how AI agents make engineers 10x faster.
They're missing the story.
Rémi Louf wrote the sharpest take I’ve read on coding agents this year. One line stopped me cold: "The bottleneck was never the code."
For 50 years we obsessed over typing speed, IDEs, frameworks — all to reduce the cost of writing code. Agents crushed that cost. What’s left exposed underneath?
People trying to agree on what to build.
Three uncomfortable truths:
→ The new bottleneck is management producing specs precise enough for agents to execute.
→ When code gets 10x cheaper, we don’t ship the same things faster. We ship 10x more things nobody asked for. Every vibe-coded product with 12 features is 11 features away from being great.
→ Context is the real commodity. Humans accrete it by osmosis — hallway chats, 2am outages, Slack threads. Agents can’t do osmosis. Whatever isn’t written down, they don’t have.
Agents are overestimated as a way to make individuals code faster. They’re underestimated as a way to make organizations externalize what they know.
Coherence was always the hard problem. Now it’s the only problem.
Building Sparkonomy as founding engineer, I see this every week; the hard part isn’t getting Claude or Cursor to write the code, it’s getting the team aligned on what the code should do. I know Priyansh Goel 10/10 relates with this 😂