AI coding agents such as Claude Code and Codex make software faster to write but do not change what software can do, according to Diogo Almeida. The founder of TypeSafe and creator of the Jev model made the case on the a16z podcast with Ben Horowitz and Martin Casado, as reported by BigGo Finance.
Almeida argues the AI industry has spent years optimizing a human judge instead of automating real work. His benchmark is blunt: if AI can solve Google-proof question answering but still cannot handle a drive-thru, the industry has been building demos, not automation.
Why AI Coding Agents Leave Software the Same
“If you use AI today to generate software, you’re still creating the same software that you did before,” Almeida said. In his framing, the tools automate software engineering without expanding what software can do.
Casado, the a16z general partner, added a data point. The average pull request at a large company is about ten lines, a figure a16z studied. “Maybe you’re writing it faster. It’s arguably getting worse just because there’s less oversight,” Casado said.
Almeida credited investor Gary Tan’s description of coding agents as “just-in-time software.” The output is still code, “maybe it’s better, maybe it’s worse, but it’s basically still code just like code looked 10 years ago.”
Jev as a Primitive Inside Code
Jev takes the opposite approach to AI coding agents. Instead of a faster engineer, Almeida wants smart software. Jev becomes a primitive included inside the code: natural language in, a state machine out, with confidence levels attached.
He is explicit that Jev is a classifier. “Classifiers are sick. Classifiers were designed to be useful,” he said. He sees AI coding agents as complementary: they handle syntax, while Jev handles semantics.
Reliability as the Moat
Almeida says reliability, not raw capability, is the real moat. “Reliability is what this thing is,” he said. He said the team could have shipped much earlier and chose not to.
He breaks reliability into four layers:
- Uptime and service-level agreements.
- Determinism, so the same functional input produces the same output.
- Robustness, meaning “similar intelligence every time.”
- Output that is smart every time, even when it is not the same function.
The SaaS Apocalypse in Reverse
When AI coding agents arrived, the “SaaS apocalypse” narrative held that cheap software would sink software companies. Almeida says software is cheap to generate but not easy to replicate well, because “a lot of the stuff happens beneath the hood.”
Companies that already own customer relationships are best placed to make their products smarter, he argues. He wants to work with “the biggest, most boring, most in-the-know” software companies. The view fits a wider debate over how founder-led companies face market tests.
He also predicts that multi-choice forms will disappear once software can map natural language to structured output on its own.
A Split View on AGI
Almeida is skeptical of recursive self-improvement. “I do not, for nuanced reasons, think we are on the path of RSI,” he said. But he calls OpenAI’s definition of AGI, automating most economically valuable work, “extremely doable.”
His case rests on how much work is rote and simple. The same AI debate runs through other leaders, including calls for government AI oversight and OpenAI’s own push into creator products.
For operators, the implication is where value may land. If Almeida is right, the winners will be companies that already own distribution and domain knowledge, and the drive-thru remains his test.