How Zillow’s engineers are actually using AI

Inside Zillow’s engineering organization, AI has stopped being a productivity shortcut and become something closer to a collaborator, one that’s present from the earliest stages of a problem through execution, evaluation and the next iteration.

The engineers building Zillow’s AI mode, deep research agents and the underlying platform that makes it all run are not just moving faster. They are working differently. The questions they are asking have changed. The problems they have time to solve have changed. And the skills that determine who gets the most out of all of it are not the ones most people would expect.

Here is what that looks like from the inside.

Six months ago, Zach Harrison didn’t trust AI tools

Zach Harrison, a senior applied scientist on Zillow’s agentic AI team, was skeptical about AI tools as recently as a year ago. He tried them, found them insufficient and set them aside. Then, about six months ago, the quality improved enough to change his mind and his workflow.

“In the past, I’d look at a problem, maybe do a literature review, come up with experiments myself and start coding them basically all myself,” Harrison said. “Now, it’s more like jumping into [AI tools] right away and using it almost like a partner — from brainstorming to literature review to coming up with a plan to bouncing ideas off of. And once that plan is in place, it’s doing most of the coding work, and I’m reviewing and testing.” 

The shift in velocity is hard to overstate. Experiments that previously took a two-week sprint now take a day. Harrison now runs three models in the time it used to take to run one.

Where Harrison has been most surprised is not in the coding. It is in communicating the work. “A big part of a scientist’s job is communicating numbers to people — and that’s maybe an underrated part of the job,” he said. “Using AI there has been quite impressive. It’s just visually better than me at presenting data, using colors, making things pop.” 

AI-generated documents still need a human touch before they go anywhere. The quality is improving, and Harrison expects that trend to continue. The more important shift is what the speed unlocks: more ideas tested, more ground covered, more time spent on the work that actually requires a scientist’s judgment.

The team that grew velocity by 50% started by rewriting its Jira tickets

Aaron Wroblewski, a senior manager of machine learning engineering on Zillow’s AI mode team, made his biggest change upstream of the code itself. His insight was simple, and its effects compounded fast: The quality of what an agent produces is determined entirely by the quality of what it is given.

“The ticket isn’t a pointer to the work,” Wroblewski said. “It is the work definition, precise enough that an automated coding agent can be assigned the task and can execute against it without further clarification.” 

When the spec is vague, the agent produces vague output. When the ticket includes acceptance criteria, technical scope and the reason behind the work, the team gets a reviewable merge request back. Using that loop, Wroblewski’s team has grown velocity by roughly 50% this year. The mechanism matters: It came from rebuilding the front end of the workflow so the spec quality could support automation, not from simply adding AI to an existing process.

That distinction shapes how Wroblewski thinks about the skill his team needs most. “The most important skill for every engineer today is communication. If you think about AI as a machine and try to program it the way you’re used to programming code, you will fail. The key difference is that AI has the opportunity to take on significant judgment and interpretation. Teaching it how to make good judgments is very different from teaching code exactly what to do,” he said. 

His team does not treat agents like tools. They treat them like new teammates. They walk them through context. They build conventions into the repo where agents can find them. They give feedback and watch behavior change. “You wouldn’t drop a new engineer into your codebase and assume they’d figure out your conventions,” Wroblewski said. “You’d onboard them.” 

The agent gets sharper every time it fails

Wroblewski’s team has taken that onboarding logic further than most. Its members have built a workflow in which the agent does not just execute work. It learns from its own failures.

When someone files a bug, an agent reads the report, cross-references the codebase and produces a structured Jira ticket with root-cause hypotheses already populated. If the fix turns out to be a prompt-engineering problem, a second skill kicks in. The agent analyzes its own failure, identifies what went wrong and proposes specific edits. A human approves the change. It ships. The next failure is different.

“We’re so close to automating the full loop — bug report, triage, fix, evals, ready merge request,” Wroblewski said. “The agent gets sharper every time it gets something wrong.” 

Humans still review code and test results before anything merges. That guardrail is not moving. The human role is shifting from writing code to writing good specs, evaluating agent output and deciding what to build next.

A weekend prototype became a full product initiative

Min Hung Shih, a software engineer building Zillow’s deep research agents, arrived at the same conclusion from a different angle. He measures what AI has changed in cycles, not hours. He kicks off benchmark runs before he goes to sleep and checks results in the morning. He prototyped a concept for AI-enhanced market reports on his own, as an experiment, and it became a full product initiative.

“I kind of feel like I’m even more of a product person right now, because there’s just a lot more space you can explore with this technology,” Shih said.

Shih has rebuilt the routine parts of his workflow to move faster so he can spend more time on the parts that require his judgment. For code review, he built a custom skill that maps component relationships visually. Instead of manually tracing how a function change ripples through a codebase, the agent does it. He uses AI tools to condense the coding style and feedback patterns of engineers he admires, and uses those learnings to gain a skill that reviews his own code in that style.

The result is a flywheel that spins faster at every stage. “It’s in coding, reviewing, how you communicate your idea, how you prototype,” Shih said. “The entire lifecycle is like a flywheel, and it’s spinning faster.”

Zillow’s institutional knowledge was stuck in wiki pages

Cody Bushey, who manages Zillow’s Agentic Data Foundations team, is solving a version of the same problem at the team level. His group sits at the intersection of platform, data and AI, and builds the shared context and memory layer that makes Zillow’s agentic systems work. His team’s biggest challenge was organizational. The operations of systems lived in people’s heads or on wiki pages that went stale the moment they were written.

His team’s answer was to turn that tribal knowledge into a reusable process. An engineer built a platform cluster upgrade skill, a lightweight, intelligent set of instructions that guides an agent or a junior engineer through a complex process, surfaces the right validations and flags the hidden blockers a wiki page would have missed.

“Before, that would be maybe a wiki page that’d go stale, or the hidden roadblock wouldn’t be surfaced,” Bushey said. “Now, we can codify that and share it across teams.”

His team also runs Slack-integrated agent bots with conversational context, so a team discussion can generate a Jira ticket and a corresponding merge request without anyone leaving the thread. Knowledge that used to dissipate after a meeting now compounds.

Bushey sees the next wave as shared context across teams. “Everyone’s creating 20-page docs. How can we use AI such that when I jump back on tomorrow, all the stuff that happened in between now and then, I’m totally caught up — accurate, high signal, not another 20-page doc I have to read every morning?”

The engineers getting the most from AI aren’t the best coders

The engineers getting the most out of AI are the ones who are best at being clear about what they want.

Wroblewski said communication is the most important engineering skill of this moment. Harrison framed it as design architecture, knowing the steps needed to reach the finish line independent of who or what is doing the coding. Bushey pointed to the capacity to go deep, be skeptical and catch hallucinations; the judgment that makes everything else trustworthy.

Each of them is describing the same shift. AI handles more of the execution. Human judgment determines whether any of it matters.

“Before, trying something new was like an undertaking — ‘Oh, this is going to take a month,’” Harrison said. “Now, it’s just: ‘I can go try this.’”

Zillow’s engineers are running more experiments, moving at speeds that were not previously available to them and thinking at a level of abstraction that the execution work used to crowd out. The ceiling on what a team can do has been raised. These are the people finding out how high it can go.