Back to What Happens When Everyone Can Build Software, with Garry Tan & Paul Copplestone.
How Do You Design Software That AI Agents Want to Use?
Paul Copplestone says you design software for AI agents by putting the fix inside the error. A generic developer-tool error, the old 500 that only says something is wrong, leaves a person to go search. An agent needs the message to say how to fix the error, or which docs to read, so the agent can remedy itself and one-shot the task.
Supabase got there by picking a use case and looping on its own error messages. He describes two stages to winning the agents. The first is being picked by the model. The second is the moment the agent gets stuck: give it the tools to fix the problem so it does not switch to something else. Agents are smart enough now to figure a lot out, and the product still has to be re-engineered from what used to be written only for humans and developers.
He was surprised, at a DevTools dinner the night before, that the idea was not already obvious. The change is not a new screen. It is the same product, rewritten so an agent can get unstuck without a person in the loop.
What Happens When Everyone Can Build Software ·
From the transcript
We'll call it the Artifactory. Oh, man, I quite like that. We talked about before this, there's just a bunch of things that you've been doing that really make agents like you better, like error messages. What was that key insight? And are there a bunch of other things that you're doing that are kind of obvious on the one hand, on the other hand, nobody's doing yet? Yeah, so like the error messages.
Yeah, I was at a DevTools dinner last night and I was surprised that this wasn't so obvious. But we really dug into our error messages so that you can take a use case like, oh, I want to build an application. And generally like developer tool error messages are like 500 something basic. And you just know there's a problem. But now you actually want the agents to like remedy themselves.