AI coding agents work mostly by reading and editing files. That creates a problem for the database behind an app. To learn what a database looks like, an agent may have to work through a long record of past changes and guess. Supabase, a company that runs hosted databases for app builders, is trying to fix this by describing the whole database in files that show its current state.
Supabase co-founders Paul Copplestone and Ant Wilson talked about this direction with swyx, host of the Latent Space podcast, in an episode published October 28, 2026. The conversation also covered the mistakes early agents made with databases and Wilson's worry that agents will push everyone to build software the same way.
From a dashboard to a folder
In the episode, Supabase's history was laid out in three stages. The company started with a dashboard, where developers managed their database through a web interface. From the start it could also be used "headless": programs could talk to it directly through an API without anyone clicking through screens, which was named as one reason Supabase succeeded. This year's big focus, the discussion said, is being code-first and Git-native, meaning the project lives in files that developers can track in version control.
When a developer starts Supabase on their own machine with its command-line tool, they get a local supabase folder that represents the project, including its database and functions. The conversation compared this to infrastructure as code, the practice of describing servers and services in text files instead of setting them up by hand.
Why a log of changes confuses agents
Supabase's databases run on Postgres, a widely used open-source database. Developers normally change a Postgres database through migrations. Each change, such as adding a table or a column, goes into a new file in a migrations folder. Over time the folder becomes a historical log. The episode called this approach "industry standard."
Agents have a harder time with it. In the episode's example, an agent asked about the shape of one table has to read through as many as a thousand migration files and guess how they add up. The alternative is to start a database locally, run every migration and then inspect the result, which was described as too slow.
Supabase's fix is a schemas folder that states the complete desired state of the database, from tables to access policies and database comments. The agent reads one current description instead of piecing the database together from its history. Agents, the discussion went, want to iterate quickly on the file system without reaching out to other systems.
The episode was recorded the day before an announcement described as Supabase's "final change" for declarative schemas, a project that had to cover all of Postgres. That announcement appears to be the one Supabase made at its Select event on October 2. According to the company's later write-up, developers edit the desired-state SQL files and a new library, pg-delta, generates the migrations needed to get there. Supabase's command-line tool loads the schema files into one temporary database and replays the migration history into another, then compares the two. Supabase says pg-delta works with any Postgres database from version 14 to 18, not just Supabase's, and is still in alpha. It generates changes to a database's structure, not to the data stored in it, and it does not handle rolling out changes to a live production system without downtime.
Images in Postgres and security turned off
Wilson remembered how badly agents handled databases at first. "Very early on, the agents were like trying to shove image data into Postgres," he said. "They were turning off RLS in order to get their projects to work."
RLS stands for row-level security. Supabase's documentation describes it as access rules enforced inside the database itself, not only in the app's screens. A typical rule compares the logged-in user's ID with the owner recorded on each row, so people can see and edit only their own data. Rules can apply separately to reading, adding and updating rows. Turning RLS off removes those per-user checks, so the database no longer limits which users can read or change which rows. An agent that does it to clear an error gets an app that works but no longer protects its users' data.
Wilson credited Lovable, a service that builds apps from plain-language requests, with tackling this first. "Lovable wrote really good skills to stop their agents doing stupid stuff," he said. Skills here are reusable instructions that steer an agent. Lovable's security guide, published in March 2026, says the company turns RLS on by default, checks for missing RLS after database changes and runs scanners over access policies, app code and dependencies. The same guide says app owners are still responsible for making sure their ownership and role rules are correct.
Wilson said the models themselves "don't do that stupid stuff anymore." He compared the progress to the Will Smith eating spaghetti meme, which contrasts crude early AI video of the actor with much more realistic later versions: it had gone "from like Will Smith eating spaghetti circa, you know, 2022 to Will Smith eating spaghetti in 2026."
Who is the customer married to?
The discussion turned to competition. App builders such as Lovable are adding their own database features, and the conversation expected companies to move into each other's territory. That raised the question of whether a company's chief product officer is more committed to Lovable or to Supabase.
Wilson answered that "it's harder to migrate your database than it is your front-end tool," the part of an app that users see. He added a qualification right away: "Not that I don't think Lovable is sticky. I think it's amazing."
Building from a phone
Asked what open questions would change how Supabase bets, the conversation focused on how development will be done and what people will build. One answer drew on personal experience: doing development entirely from a mobile phone, which was described as comfortable. The caveat was that this works for someone who knows what is happening behind the scenes and can tell the agent what the infrastructure needs.
The forecast went further. A person with no engineering background might tell a phone or personal agent "I want to build X" and have a live application one second later, then keep changing it themselves. That raised questions Supabase is still working through. Will such apps be personalized? Will they need to scale up? What tools does Supabase have to provide to stay the default provider?
Wilson's worry: one way to build everything
Wilson raised a concern of his own. After a few decades as a software engineer, he said, he has strong opinions about how software should be built, on practices such as continuous integration and continuous delivery, which test and release code automatically. He catches himself pushing those opinions onto agents, although he increasingly doesn't need to.
"My worry there is are they all just going to like converge on one way of building software," he said. He pointed to the industry's "famous flame wars" over the correct approach and asked whether those would disappear, leaving software stuck at what he called a "local maxima": a design that looks best from where you stand but is not the best one available. The discussion gave this the machine-learning name mode collapse, used for a model that keeps producing a narrow range of outputs, along with a joke that it would be great if they all used Supabase.
The response in the conversation called it a valid fear, and one outside Wilson's control. Casual app builders, the discussion went, will converge on whatever the model's default is, which was identified as Supabase, while discerning developers working at high scale will always care about the details, and their choices would eventually filter down.
Wilson asked whether those highly skilled people would still be trained. The reply began with "I don't know." It went on to suggest that the average college may be in trouble because people can teach themselves. The top tier, it said, has always been driven by curiosity, like the programmers who take time off to attend the Recurse Center and work on interesting programming problems instead of earning industry salaries. The conversation spoke of taking a year off. The Recurse Center describes itself as a free, self-directed programming retreat whose batches last six or twelve weeks, in New York or online.