Even in the areas made vastly more efficient by today’s agents, there’s still room for judgment.

David WoollardDavid WoollardCTO @ Standard AI

AI agents now handle much of the work of writing code, and engineering leaders are rethinking what they need from the people on their teams. Deciding what to build and which results to trust still falls to experienced engineers, and that skill is becoming a hiring priority. The companies getting the most out of AI also tend to be well organized, with code and documentation clear enough that a smaller, more efficient model can do the job.

David Woollard is Chief Technology Officer at Standard AI, which turns retailers’ existing camera infrastructure into data on how people move and behave in physical spaces. He joined the company in 2019, during its autonomous checkout era, and has helped lead its technology through each phase since, after running engineering at several Bay Area startups. Woollard holds a PhD in computer science from the University of Southern California and regularly publishes AI experiments that test how models reason and how to measure them. His own team, like most, is still writing its rule book for AI adoption.

“Even in the areas made vastly more efficient by today’s agents, there’s still room for judgment,” Woollard says. Agents can produce working software quickly, but someone has to decide whether the architecture holds up and whether the output is safe to ship. How well a team handles those decisions also shows up in what kind of model it needs.

Right-sizing the model

For a while, many engineering teams defaulted to the most powerful model available and absorbed the token bill. More compute doesn’t always mean better results. In one test with Mistral Large 4, the newest flagship model from French AI company Mistral, Woollard found that turning on the model’s reasoning mode more than doubled token use with no measurable gain in accuracy. “We’re out of this token-maxing phase of AI. I’m seeing a lot more conversations about lighter-weight, more efficient models, whether that’s from a cost perspective or an operational perspective,” he says.

Costs also climb when a team hands an agent a loose request and stops watching it. “You might have started with a fairly simple request to a coding agent, didn’t babysit that request well enough, didn’t put enough controls in place, and you arrive at something that should have been a 20-line script. It ended up with a thousand unit tests and a lot of verification, but might not have even solved the original problem,” he notes. Scoping a request clearly and checking in along the way keeps an agent focused on the problem it was given.

One useful test of AI readiness is how powerful a model a team needs to get its work done. Teams that get good results from smaller models tend to have more mature engineering organizations. “If you’ve done a good job making your code base accessible, readable and parsable, your knowledge base is in sync with your code, and your general workflows are well understood and documented, then you might not need whatever the latest and greatest frontier model is. It shifts away from fire-and-forget behaviors and toward more responsible stewardship of the tooling,” Woollard says.

Setup matters too. Agents work best with testing and review built around them and a clear point where a person signs off. Those steps affect what AI costs as much as the model does.

Humans in the loop

AI helps every kind of technical team, but in different ways. Its biggest impact depends on what was slowing the team down before. “I have research scientists who, in the past, would have been paired with developers building out infrastructure around their research. Now those folks can be more self-sufficient, and for a lot of them, it’s felt very empowering,” Woollard says.

That also frees developers to spend more of their time getting work ready for production.

Agents can take on a growing share of the work, but people set the direction. “There are really important decisions that get made in the process of developing even very well-known patterns of software. There’s still important work for humans in the loop there,” he says.

Experienced engineers guide and review what agents produce, and that skill takes time to build. Companies that keep hiring newer engineers give them room to learn it.

Wael Sabra is Founder and CEO of SiiRA, which helps companies hire remote engineers and AI teams. Before starting SiiRA, he spent eight years building data and ad tech products at Cablevision and Altice USA. He sees a particular advantage in junior hires. “AI adoption is a change management challenge first. Juniors come in without years of ‘this is how we’ve always done it,’ so they’re often open to new ways of working. Pairing them with experienced engineers gives a team both perspectives,” Sabra says.

There’s plenty of room in technology for new graduates. Technical skills go even further when they’re paired with knowledge of another field, like economics, marketing or design.

Wael SabraWael SabraCEO @ SiiRA

Cleaning house

AI is giving companies a new reason to fix old problems in their code and processes. “It’s as solid a forcing function as we have so far. Engineering debt is compounding in agentic complexity and token costs, and it’s forcing a spotlight on the discipline that’s necessary to be a mature organization in this day and age,” Woollard says.

Larger, older companies usually have the most to fix, after years of systems built up by different teams.

Sabra sees the same gap from the hiring side. “Much of what companies want to hand to AI already exists as a workflow. The trouble is that it often lives only in someone’s head. Once a team writes it down, an agent can run it and a new engineer can pick it up, and the company stops losing that knowledge every time someone leaves,” Sabra says.

Woollard compares the cleanup to preparing a company for a security audit. “If I walked into an organization of 20 people one year into its existence and said we’re going to get SOC 2 compliant, that’s an easier hill to climb in some ways than walking into a 10-year-old organization with 150 people doing their own thing,” he says.

Hiring for curiosity

The qualities that make a good engineer haven’t changed much with AI. Solid fundamentals and the ability to adapt matter more than expertise in any single tool, because tools and problems keep changing. Curiosity and a scientific mindset help too. “I like to have a hypothesis, and I like to prove it out in some way, shape or form. You follow the evidence, good, bad and otherwise, and figure out what to do next,” Woollard says.

For people entering the field, software development remains a strong career path, especially for those drawn to the work itself.

Sabra, who started his career as a software engineer at UBS, has advice for students deciding what to study. “There’s plenty of room in technology for new graduates. Technical skills go even further when they’re paired with knowledge of another field, like economics, marketing or design. My own career grew out of building data products for advertising, where understanding the business made the technology more useful,” Sabra says.

Curiosity can be taught, and AI makes it easier to practice. Woollard and his daughter play a game while they wait for the school bus. She asks whatever question is on her mind, and they spend a few minutes exploring it with an AI model. Their topics have ranged from animal taxonomy to 17th-century timber framing.

For engineering leaders, that makes curiosity a quality to hire for and keep encouraging on the job. “Whether it’s for your own career or learning about a new topic, all of those things are possible in ways they’ve never been possible before. If you’re willing to run headlong into that change, it’s an exciting time,” Woollard says.