Treasure your fuck ups

Bite into a Kit Kat, and you’re eating failure. Between each perfect wafer is a layer of ground up rejects. That’s exactly how we should build agents.

Since 2022, I've built a lot of shitty AI solutions. A few absolute bangers. But also a mound of clever ideas that just didn't work. And that's fine. Because every failed agent should feed a successful one. And not in some kumbaya "we're a learning organization" way, but by concretely capturing the learnings, codifying the decisions, and making them public. Every experiment should be ground up to build the layers of common law.

Most enterprise gen AI work skips that part. A solution ships. A team disbands. And the policy calls, design principles, data governance arguments, and prompt that finally worked on the eighth rewrite all end up in a dusty folder, a dead Slack channel, a forgotten repo. The next team starts from scratch. Same arguments, same mistakes, same six-month ramp.

Killed projects are even worse. Months of expensive work is vaporized the instant leaders pull the plug.

McKinsey puts a number on the cost: Companies that build reusable assets (think preapproved prompts, application patterns, governance baked into the platform) eliminate 30 to 50% of the nonessential work on every subsequent project. Approval cycles shrink by up to 90%. The savings compound.

Every project (even the failures) should seed the next one. Markdown files. Decision logs. Governance patterns. Crushed up and folded into whatever comes next.

Everyone craves the chocolate.
But the failure holds it together.

—
Source: “Overcoming two issues that are sinking gen AI programs,” Baig et al, June 2025