BUILT BY YAHOO

A Human Path to Agentic Software Development

The way software gets built has changed, and it keeps changing fast. Agents can now write, test and review code, not just suggest it. But simply giving teams access to the tools isn’t enough. Agents need code that’s organized for them, with checks for security, privacy and quality built in. And people need to learn differently, too: by experimenting, asking questions and getting comfortable with what doesn’t work the first time.

That’s why Yahoo built its Agentic Development Program, or ADP, to transform how software is built at the company. It’s a more human path to AI transformation, pairing new technology and ways of working with a significant investment in the people making the shift. And it’s applied directly to existing products instead of just using agentic development for greenfield projects.

Early results show what that makes possible. Projects traditionally measured in months are getting done in weeks, without sacrificing quality, and teams are using that speed to try more ideas and get products in front of customers sooner.

“Building has changed, and the companies that lead will be the ones that help their people change with it,” says Yahoo President and COO Matt Sanchez. “ADP gives every team the foundation, the time and the support to work this way on everything we build.”

Building has changed, and the companies that lead will be the ones that help their people change with it.

How Yahoo built the program

The first step in ADP is standardizing on a common set of agent-ready tools across Yahoo, from build-and-release pipelines to monitoring and ticket tracking, with security built in. Over the years, Yahoo’s engineering had become fragmented, spread across tens of thousands of code repositories and a patchwork of legacy systems. People can work around that, but agents need consistent tools and well-organized code to do good work.

Teams refactor and consolidate their code to adopt the common tools. With agents doing much of the work, they can pay down technical debt that once would have been too time consuming and costly to tackle. Individual teams still decide how they work day to day, but everyone builds on the same base.

Working with agents also requires new skills and habits. Yahoo tested how to teach product teams through two rounds of pilots before rolling out ADP more broadly. The first showed that teams needed more context before they arrived and more hands-on building once they did. So the second moved introductory training ahead of the workshops, rebuilt the in-person sessions around hands-on work and gave every team dedicated Change Champions, colleagues who already work this way. The program has kept evolving with each wave since, incorporating feedback from participants in real time, iterating upon every session to improve both the experience and the outcomes.

And instead of a login and a how-to video, ADP brings whole teams together in person, in Yahoo offices around the world, to rethink familiar ways of working. Asking basic questions and trying things that don’t work are part of the process. Change Champions explain the why. Coaches from Zartis, a firm that specializes in agentic development, help with the how. Going through the program together also gives everyone, whatever their role, a shared picture of what working with agents should look like. And coaching continues long after the workshops through a three-week, high-support build phase with embedded Change Champions and office hours with Zartis technical coaches.

The real test is what teams build with it, and three very different teams are already showing what’s possible.

Bringing a critical ad capability in-house, in 12 days

Yahoo has long relied on an outside partner for a critical piece of how ads are valued on its properties, at a significant annual cost. On Yahoo's Monetization Signals squad, two engineers and a product manager set out to bring that capability in-house.

They estimated the work at 35 days. Working agent-first, they shipped it in 12. None of the code was written by hand. The engineers directed agents and reviewed what came back, and people signed off on the design, privacy safeguards and production changes. The feature went live with zero rollbacks and gives Yahoo a measured path to reducing third-party costs.

“The speed is what people notice, but what made this possible was building agent-first on shared tools and a shared codebase,” says Amit Agarwal, a distinguished software development engineer on the team. “That’s what let a small team take on something like this at all.”

The speed is what people notice, but what made this possible was building agent-first on shared tools and a shared codebase.

Agent factories: a yearlong build in three weeks

The small-business advertising market is dominated by a few giants. The Yahoo DSP team saw an opening, but a mobile product for small businesses would traditionally take six to 12 months to build. With little immediate revenue at stake, it was the kind of project that usually gets pushed down the roadmap. A six-person squad delivered a working mobile web app, from sign-up to AI-generated ads with voiceover, in three weeks.

The key was agent factories: AI agents that pick up tasks from a shared backlog, write and test the code, and hand it to engineers for review. With several running at once, the team shipped 730 code changes in three weeks, 136 of them in a single day, backed by more than 4,100 automated tests. No critical defects reached production.

"We didn't spend our time writing code. We spent it trying ideas, throwing out what didn't work and trying again," says John McClean, VP, Engineering, Yahoo DSP. "With agents writing the code, we could rebuild the product as many times as it took to get it right."

Alpha testers are now lined up, months ahead of plan.

Security experts build a new line of defense

As AI helps attackers move faster than ever, Yahoo's security team, the Paranoids, is moving to find and fix vulnerabilities faster, too.

A team of incident responders, bug bounty program managers and product security leads built a tool that tests Yahoo’s defenses the way an AI-powered attacker would. The agentic tool links small weaknesses to reveal bigger ones and flags them for fixing, giving Yahoo’s security experts a clearer view of where to focus. Working part time, they built it in just over a month. Everyone on the team committed working code from day one, and a new idea could go from idea to code ready for review in under 20 minutes.

Traditionally, red teaming, where security experts play the attacker, is resource intensive: teams can spend months preparing for operations that last only a few days, so most companies can run just a few a year. The new tool gives Yahoo’s security experts a way to test far more often.

“Most of the people who built this tool don’t write code every day,” says Chris Keast, VP, Paranoids Cyber Defense. “With agents handling the implementation, they turned what they know about how attackers think into software that adds another layer of protection for Yahoo’s systems and the people who use them.”

Key takeaways

Start with common, agent-ready tools. Standardizing the systems teams use to build, test and ship, and organizing code so agents can work with it, is the critical first step. This lets teams build on existing production systems with legacy codebases at agent speed.

Invest in people. New technology only goes so far without the time, coaching, and support to use it well. Bringing whole teams through the change together builds the skills and confidence that make it stick.

Keep people where judgment matters most. Let agents handle more of the execution, and keep people in charge of design, quality, and key decisions. Speed and safety can grow together when guardrails are built into the system.

To learn more about overcoming barriers to AI transformation, check out this recent Newsweek webinar featuring Yahoo President and COO Matt Sanchez and other technology leaders.

‍