How to build an AI native product for the agentic era

Ivan Burazin at Daytona has a simple rule for how to build an AI native product: design for agents as the primary consumer. Jon Noronha at Gamma described the human side, using AI to drop new users into a fully worked-out example instead of a blank page.

How do enterprises deploy agents in production?

Justin Reock, Deputy CTO at DX, shared several examples. Spotify deployed an agent that pulls context and remediation steps from runbooks into SRE Slack channels when an incident is detected. “Those critical few minutes of discovery that have to happen are now being done by an agent,” he said. “They're doing about 90% of their incidents through this now, and it's significantly reduced their MTTR.”

At Zapier, engineers write an agent for any task they've done twice, such as summarizing commit logs. “They brought their standups from five times a week down to just two times a week now doing that.”

Morgan Stanley used agents to create specs from legacy code instead of reverse engineering it by hand. Reock estimated, “they're saving over 280,000 hours a year just by eliminating that part of the reverse engineering step.”

Why does AI-assisted engineering sometimes reduce productivity?

Reock cited a study where “every engineer who took part in this study felt like they were more productive. They self-reported that they thought they were getting more done, and actually, the data showed there was about a 19% overall reduction.”

A 2% increase in change failure rate against a 4% industry benchmark means shipping as many as 50% more defects than before.

Asked whether AI writing tests for its own code would just write passing tests, Reock said, “Absolutely, there is that concern, which is why we still need humans in the loop validating this stuff.”

How to build an AI native product when the agent is the user

Burazin was direct on a panel: “the agent is the end consumer.” He warned, “If you are not, the category you're building from will not be relevant in the future.”

Reock agreed that documentation will shift toward agents: “I think we will be living in a world where the agent documentation and markdown is way more important.” He described a front-end React developer who built an MCP server to pull Chrome developer tool screenshots and feed them to the model, giving it visual context for what broke.

How do generative tools rewire user onboarding?

Noronha said, “the very first thing that AI let us do was overcome that blank page. We could drop you into a fully worked-out example based on your prompt.”

He argued that taste and design quality are a core differentiator for Gamma, and that the team's job is to give non-designers great taste. Andrey Khusid at Miro said taste and product quality matter more than ever because Anthropic, ChatGPT and Google keep raising the bar.

What architectural mistakes do early AI teams make?

Adam Cheyer, co-founder of Siri, admitted the Siri team got it backwards: “But we weren't very good at knowing. We force fit knowing into a doing architecture.” He sees the reverse mistake now: “a lot of the agentic work I see is trying to force fit doing into a knowing framework. It's slow.”

Cheyer also recalled that in 2009, Sentient used two million CPUs and GPUs for distributed computation to build the first large-scale machine learning platform.

Put together, the advice on how to build an AI native product is consistent: build for agents as users, use AI to remove the blank page for humans, and keep humans checking the output.

To hear more from the people putting AI into production, see the HumanX 2027 program preview.