Code Review Is the Real Bottleneck in the AI Software Factory

The speaker in this clip is Dexter Horthy, CEO and co-founder of HumanLayer, shown at the end as a keynote speaker for AGNTCon + MCPCon Europe in Amsterdam. His argument in one sentence: agents make implementation fast, but review, testing, and verification set the real pace of an AI software factory, and you should optimize within that constraint instead of pretending it does not exist. The video opens with an on-screen hook that reads "SWEs are getting lazy and gaslighting themselves."

The Software Factory Framing

Horthy starts from a simple substitution model. Take the pre-AI software factory and swap the human who builds the thing for an agent that builds the thing. Implementation that used to take days now takes minutes or hours. The on-screen visuals reinforce the framing with a robot typing at a workstation and an issue tracker full of tasks, bugs, and epics being routed into the pipeline.

Why Faster Implementation Does Not Mean Faster Shipping

The tempting conclusion is that the whole pipeline just got faster. Horthy pushes back: if you still want to test and review the same amount of work, that part of the pipeline has not sped up at all. Even if you add agentic code review and agentic testing, and even if you route every incident and feature request into the factory, two limits remain:

1. How much work you can stuff into the queue
2. How fast you can review each change that comes out

The second limit is the one people ignore.

The "Let the Model Cook" Failure Mode

Horthy describes a specific rationalization he sees engineers make. The review step is slow and unpleasant, so they convince themselves the models are smart enough to skip it. The internal monologue goes: "I'm going to let the model cook. For better or for worse, this is how we do it. I accept the risks." He calls this gaslighting yourself. The models being impressive does not make unreviewed code safe to ship.

Accept the Constraint, Then Go Fast Inside It

His prescription is to treat review throughput as a hard constraint rather than a nuisance. The questions that define your factory's speed are: how fast can you review a change, understand it, verify it, know it is good, and know it preserves maintainability. Accept that constraint and then push to go as fast as possible within its bounds. Speed gains come from making review faster, not from deleting it.

Context: Who Is Speaking and Why

The video is a promotional clip for AGNTCon + MCPCon Europe, an Agentic AI Foundation event in Amsterdam on September 17-18, 2026, aimed at AI builders, founders, and innovators. Horthy keynotes the event. HumanLayer builds tooling for human oversight of AI agents, which explains why review-as-bottleneck is his core talking point. The framing serves his product thesis, but the argument stands on its own.

Key Takeaways

• Replacing human implementation with agents speeds up only the build step, not testing or review.
• Agentic code review and agentic testing still leave you bottlenecked by queue capacity and human review speed.
• Skipping review because "the models are smart enough" is a rationalization, not a strategy.
• Treat review, verification, and maintainability checks as a fixed constraint and optimize speed within it.
• The teams that win will be the ones that make review faster, not the ones that remove it.

Resources

• HumanLayer -- Dexter Horthy's company, building human-in-the-loop tooling for AI agents
• Agentic AI Foundation -- organizer of AGNTCon + MCPCon Europe, Amsterdam, September 17-18, 2026 (URL shown on screen in the video)