Diogo Almeida warns coding agents fail at application logic

Software development economics broke overnight, but the code isn't getting any smarter.

On This Week in Startups, Jason Calacanis proved how cheap software creation has become by placing a $5,000 bounty on a custom Chrome extension. Twenty years ago, building the web-annotation tool would have cost $500,000 and taken six months. Within two weeks of posting the challenge in late September 2026, developers submitted three working prototypes using vibe coding. Developer Robert built annotated.wtf, an app that clips videos and embeds source media without standard engineering overhead.

Yet beneath the surface of these rapid prototypes lies a structural ceiling. On The a16z Show, TypeSafe founder Diogo Almeida cautioned that tools like Claude Code and OpenAI's Codex merely act as faster typists. They spit out thousands of lines of imperative syntax per minute while leaving underlying application logic completely untouched.

Investors see this friction play out in real time. Calacanis noted that while vibe coding lets developers build rapidly, demo quality still dictates valuation. Submission creator Chirag Azarpoda built a working browser extension for the bounty, but clunky user interfaces and manual timestamp entry weighed down its performance. Speed in code generation means little if the end-user experience stumbles over basic workflows.

The glut of rapidly generated code presents deeper systemic risks. On The a16z Show, venture partner Ben Horowitz warned that autonomous coding agents threaten to make software worse by generating unmonitored and insecure code at scale. Rather than relying on external AI wrappers, Almeida argues developers need embedded decision nodes inside their state machines to yield deterministic behavior.

TypeSafe designed its classifier tool, Jev, to embed probabilistic intelligence directly into existing state machines. Almeida rejects the industry belief that a single massive frontier model will govern all computing tasks. Instead, software engineers require granular control over non-deterministic outputs within defined execution budgets.

The economics of software are shifting from raw creation toward data licensing and deep workflow integration. Calacanis plans to incubate top bounty submissions directly inside his firm, viewing crowdsourced builds as cheap venture deal flow. Beyond basic subscription fees, these tools capture structured human annotation data that can be licensed to train frontier language models.

Wall Street originally feared vibe coding would render enterprise software obsolete, but incumbents retain the upper hand. Enterprise platforms sit on proprietary customer data and complex workflow integrations that simple code generators cannot replicate. The real value lies in embedding intelligent primitives directly into established enterprise software.