Coding agents can generate thousands of lines of code in minutes. Problem: Most of it can’t be deployed. It breaks internal standards, fails compliance checks, or creates more cleanup work than savings.
"You can generate a lot of code, but it doesn’t really mean anything, right? It should be code that’s integrated, that’s compliant, and you don’t want to do more work on the back end, just because you’ve sped up the code generation process on the front end," said EY Global CTO Engineering Leader Stephen Newman.
EY’s product development team solved this by connecting coding agents to their engineering standards, code repositories, and compliance frameworks. The result: 4x to 5x productivity gains across the teams building EY’s audit, tax, and finance platforms.
But the benefits didn’t just come from turning on a device. Newman’s team spent 18 to 24 months building the cultural foundation and technical integration that made large-scale semi-autonomous coding work possible.
The first step was cultural. EY started with a GitHub Copilot-style tool, helping engineers get comfortable with rapid engineering and assistive AI. Newman said the key lesson is that AI should be adopted organically rather than imposed by leadership. "It is important to have organic adoption of AI capabilities at the grassroots level rather than imposing them on users." He said.
Developers wanted to move beyond code generation to build, deployment, and operations. But without deeper integration, productivity growth remained stagnant.
Newman realized that agents needed access to EY’s code repo, engineering standards, and source catalog to generate deployable code. without him "reference universe," As Newman says, agents produce generic outputs that require extensive rework.
EY evaluated several agent platforms: Lovable, Replit, and Factory’s IDE-based Droids. Rather than mandating any tool, Newman’s team measured adoption, usage, and productivity across all three.
"We as a leadership team didn’t want to be too prescriptive about identifying and decommissioning a device," Newman said. developers "actually navigated gravity and" For the factory, that became a signal that it provided real value.
factory adoption "flared up like wildfire" Once promoted from appraisal to pilot. EY had to throttle traffic to Factory and Droids and restrict which repos could connect before receiving compliance and security sign-off.
workload classification framework
The enthusiasm of the developers made it clear that EY needed discipline in delegating assignments to agents. Newman’s team divided the tasks into two categories:
high-autonomy tasks Agents handle well:
- code review
-
documentation
-
correct the fault
-
Greenfield Features
complex tasks Which still requires human oversight:
- massive refactor
-
architectural decisions
-
cross-system integration
EY also changed developer roles. Instead of writing all the code themselves, engineers became orchestrators directing agents to the right databases and repos.
With security guardrails in place and integration into code repositories completed, EY measured efficiency gains ranging from 15% to 60% across different individuals in the early adoption phase.
"We’ve taken a leap forward on many of our products, where we have what we call horizon model development, where we have semi-autonomous agent execution at scale, a team of orchestrators as opposed to doers and we have integration into the context universe," Newman said.
Newman acknowledged that it is difficult to attribute the 4x to 5x productivity gains solely to coding agents. Improvements came through trial and error along with cultural and behavioral changes in developer teams.
<a href