Skip to content
The AI-First Web Talking point

Supabase's CEO says agent-led buying turns on being the default and being fixable

Supabase's CEO says winning AI agents means getting picked by the model, then helping it recover when stuck.

W
WebPulse Newsroom
AI-assisted · 2 min read
Share on X LinkedIn
Supabase's CEO says agent-led buying turns on being the default and being fixable
In brief
  • Paul Copplestone of Supabase said winning AI agents has two stages: being picked by the model, then not being dropped when the agent gets stuck.
  • His company rewrote its error messages to tell agents how to fix problems, and tests its product against many models.

Paul Copplestone, co-founder and CEO of Supabase, said that earning an AI agent's loyalty takes two steps. He spoke on the Lightcone, Y Combinator's show. A model must first choose your product. After that, your product must help the agent recover from trouble, so the agent stays instead of moving to a rival.

What was said

Supabase sells database services. The hosts said the company has 13 million users and creates over a million databases a week. One host put agents at something like 70% of new activity.

Copplestone said the company moved from a dashboard company to an API company, and now to a code-first one. Agents, he said, mostly want to use command-line tools. He also said being chosen as a default database inside the language models helped a great deal, partly through timing.

His example was error messages. Developer tools often return a bare 500 code, which tells a person almost nothing. Supabase now puts the fix, or a pointer to the right docs, inside the message. As he put it:

"there's like two stages to winning the agents the first is being picked by the model and then the second is like if they get stuck you don't want them to switch away"

He added that Supabase tests its product against many models and keeps its own evaluation suite. A lot of its engineering effort now goes into testing specific use cases rather than building.

Why it matters

Our reading: if an agent does the choosing, the sales deck may matter less than three practical things. Does the model name you? Does your command-line tool work without a person watching? When it fails, does the message say what to do next?

For buyers and engineering leaders, error text becomes part of the product. On Copplestone's logic, a vendor whose failures end in a dead stop risks losing to one whose failures end in a fix. A host also suggested that agents like open source because they can read the code when something breaks. That would make openness part of the pitch.

The other side

This is one vendor describing its own growth. Nothing in the excerpts shows agents actually dropping vendors over poor error text. Copplestone was candid about other limits. On open source, he said it is "hard to know the counterfactual", though he thought it was critical. He also credited the timing of when Supabase became a default choice.

The conversation does not explain how a model decides which product to pick. A newcomer would not learn from this episode how to get chosen in the first place.

Copplestone also said the fix is "not actually that complicated" because agents are smart enough now. Even so, he said a lot of what was built for humans must be re-engineered for agents.

Written by the WebPulse Newsroom with AI assistance, and checked by our editorial review: every quotation was verified against the recording's transcript. How we use AI.

The conversation this talking point comes from