Domain expertise emerges as key asset in agentic AI era

Technology leader Aaron Brethorst has published an opinion piece arguing that domain expertise is the primary competitive advantage in the age of agentic AI. In his article, titled "Domain Expertise Has Always Been the Real Moat," Brethorst contends that agentic AI has severed the traditional link between domain understanding and software production, shifting the professional bottleneck from building systems to verifying their correctness.

Brethorst, a Seattle-based technology leader and transit enthusiast, notes that the hard part of writing software has never been the coding itself, but rather building a working model of the domain in one’s head. He illustrates this with examples such as payroll systems, where understanding garnishments and pre-tax deductions is crucial, and transit apps, which require knowledge of GTFS feeds and the distinction between trips and routes.

The article suggests that agentic tools have collapsed the traditional career ladder where engineers slowly learned domain knowledge through shadowing and production errors. Brethorst contrasts two archetypes: a domain expert with no software background, such as a logistics dispatcher or clinical coder, who can instantly identify incorrect AI outputs, and a strong generalist engineer who cannot distinguish plausible-looking wrong answers from right ones in unfamiliar domains.

He argues that the domain expert holds a competitive advantage because they possess the 'ground truth' required to validate AI-generated outputs. While the mechanical skill of turning ideas into code has become less valuable, the ability to judge the correctness of outputs has become the scarce resource. Brethorst advises engineers to acquire deep, tacit knowledge of specific industries or regulatory regimes, as this expertise remains scarce and cannot be replicated by AI.

This perspective marks a shift from Brethorst’s previous stance in late 2025 or early 2026, where he held the "standard take" that AI tools amplify senior developers because they possess judgment. The new argument posits that the binding constraint has moved from whether one can build a system to whether one can tell whether it is right.

Brethorst advises that the most valuable person in this new landscape is one who can verify at both layers: knowing the generated code is sound and knowing the answers it produces are true. He urges experienced engineers to bet on acquiring a deep, verified model of a real domain, whether it be an industry, instrument, or physical process, as that is the part agents cannot do for them.

The piece reflects a broader industry conversation about the impact of agentic AI on software engineering roles, specifically the shift from code generation to code verification. Brethorst’s analysis suggests that the ability to produce software without building the underlying domain model breaks long-held assumptions about how the profession is organized.