Coding agents have a habit that causes trouble for the companies whose tools they use: they write code the way it appeared in their training data. Erik Bernhardsson, CEO of Modal, said agents "love using, you know, old versions of modal syntax," and that "it's kind of hard to nudge them away from that." Because of agent training data, he said, "it's very hard to deprecate things now." Over the last year or so, Modal has become "a little bit more conservative" about deprecating and removing features.
Bernhardsson, who founded Modal, discussed this on Latent Space, in an episode published October 22, 2026. According to the publisher's description, the interview followed Modal's $750 million Series D at a $16 billion valuation. Modal sells AI infrastructure. Customers use its software development kit (SDK), a code library for controlling the platform, to run inference, training jobs and sandboxes (isolated environments where code can run safely). Modal manages the computing capacity underneath. Increasingly, the code written against that SDK comes from agents.
Good for developers, good for agents
Modal has long put developer experience at the center of its pitch, so the conversation turned to whether it still matters when agents write the code. Bernhardsson said it does. In practice, he said, "almost all the things that make developer experience good also makes the agent experience good": good documentation, descriptive error messages and fast feedback loops.
Some things have changed, he said. He recently had to set something up with Cloudflare, the web infrastructure company, which he finds "notoriously annoying to work with sometimes." He gave the job to an agent, and it configured what he needed in Pulumi, a tool that defines cloud setups in code. Agents make it easier to work with tools that used to be painful, he said, and "the sort of human toll of dealing with YAML has arguably gone down a little bit." YAML is the text format behind many configuration files.
He did not take that as a sign that the interface matters less. He said infrastructure over time "becomes and needs to be more programmatic." Modal's thesis, especially for sandboxes, is that anything once configured in YAML should be an SDK. The difference is between declaring a desired setup in a special-purpose configuration language and writing ordinary code as steps: launch a sandbox, do something with it, tear it down. In his view, SDKs "are as relevant as they always were, if not more."
Why old features are hard to retire
Asked whether Modal had changed anything in its product because so many users run agents on it, Bernhardsson said yes. The company runs many internal tests in which it tells agents to do jobs such as building a mobile app and running it. The "very concrete thing" he pointed to, which he called "kind of funny," was the trouble with deprecation.
The fix is laborious. Before Modal removes an old way of doing something, he said, "we have to be very careful to update every single example that's out there online." Skills, the packaged instructions developers can give an agent, came up as one remedy, and Bernhardsson agreed they help. He added that there is also "an insane amount of" Modal code online now that feeds into training data.
Checking whether agents recommend Modal
The second change is a version of search engine optimization aimed at agents. Bernhardsson said Modal runs internal benchmarks in which it asks an agent to deploy Stable Diffusion, the open image-generation model, using whatever tool it likes, and checks whether the agent recommends Modal. Other prompts ask an agent to spin up some sandboxes or to rank all the sandbox vendors. Modal tracks the results over time.
The question matters more when agents choose the tools. In late September, a discussion on Moonshots about building for agents included the observation that coding agents asked which tool to use had gone from recommending three options to recommending one. Modal's benchmarks are a vendor's attempt to measure that from the inside.
Asked what Modal does when an agent doesn't pick it, Bernhardsson said that part is harder. As with search engine optimization, he said, it is "primarily through content. It's, like, you just have to put out a lot of content."
The tests also show how agents perceive the company. Modal can ask an agent to rank sandbox vendors by reliability or by price and see what comes back. Bernhardsson said agents seem "a little bit more prone to" misunderstanding pricing. So Modal spends a lot of time making its content clearer and easier for agents to understand, "because we know they're all indexing, like, everything that's out there."
The slop question
The next question was the obvious follow-up: had Modal thought about creating a lot of GitHub repositories and having agents fill them with Modal code? Bernhardsson said it is "probably something you can do." His guess is that the data used to train today's language models is "quite insensitive" to whether content was written by AI, so a company could flood it with generated code. "We thought about it," he said. "Like, I don't love that idea." He did not say Modal had done it.
He compared it to search engine optimization in the early days, when auto-generated content was "a very effective strategy," and said that to some extent it still is.