AI agents got easier to build, but better results still take deeper work

Alexander Shevchenko of Ramp Labs said on the AI Engineer show that building an AI agent is easier in 2026 than in 2023. He still believes teams can dig a layer down and give users better experiences. That work, he said, takes a bit more effort now.

What was said

Shevchenko compared the two years. In 2023, with GPT-4, an agent needed very strict guardrails, meaning rules that limit what it may do. Models could not backtrack, so a wrong turn went unnoticed. Shevchenko said "some synonyms could change the behavior of your agent."

By 2026, he said, models are much better. You can give them tools and wide access to context, and they understand the task well.

His example was a spreadsheet agent at Ramp. Users drop in many spreadsheets and ask for a financial model. Opus 4.6 first made a plan, then read all the sheets. It was slow, and its thinking tokens were costly.

So Ramp trained a smaller model that answers finance questions. A frontier model still plans the job. When it needs facts from the sheets, it asks the fast, narrow model. That takes the slow, costly reading off the frontier model's job.

Shevchenko said Ramp's model beat the frontier model on exact task accuracy. In his words, "we actually got above Opus 4.6, which was the frontier back then". He added that it also ran at Haiku-level speed.

Why it matters

In 2023 the skill was careful wording and tight guardrails. Shevchenko's 2026 account describes a different order of work. His team first looked at where a strong general model fell short. Only then did it build a narrower model for that gap.

The judging came first, and the building followed. The building was still real work: Ramp trained and shipped the specialist model. But someone had to see that planning over many spreadsheets was the slow, costly step.

Our reading: the scarce skill is knowing where to build. For managers, that suggests asking where a team's effort now goes. Prompt tinkering looks like the part that is shrinking. Knowing the domain well enough to spot a model's weak points looks like the part that is growing.

The other side

Shevchenko did not say that effort moved or that judging is the new skill. Those framings are ours. His own example is mostly making: Ramp trained a model and shipped it. The facts support a narrower claim than "judging replaces making."

He called the deeper work only "a bit more effort." He gave no figure for its cost or the staff it needs.

Ramp is an application company with its own applied research group. Many buyers of software have no such team. The excerpts do not say how a smaller firm would find such gaps. They also do not say how people would learn the skill. The accuracy result is his own account of one task at Ramp, not an independent test.