Google Engineer Stunned as AI Coding Tool Recreates a Year’s Work in Just One Hour
 |
| Google Principal Engineer Jaana Dogan, who works on the Gemini API, said on X that a competing AI tool replicated a system design that her team had constructed during the previous 12 months. |
Google’s Jaana Dogan shares a firsthand experience testing an AI coding tool that rapidly matched a complex, year-long engineering effort.
For years, the race to build smarter AI coding tools has been framed as incremental—small boosts in productivity, cleaner code, faster debugging. But a recent experiment by a senior Google engineer suggests something far more disruptive may already be happening.
The speed at which artificial intelligence is reshaping software development has surprised even the people building it. Jaana Dogan, a principal engineer at Google working closely on the Gemini API, recently shared an experience that left much of the tech community rethinking long-held assumptions about how quickly complex engineering problems can be solved with the help of AI.
In a widely discussed post on X, Dogan revealed that she tested Anthropic’s Claude Code—an AI-powered coding assistant—by giving it a high-level problem description. What came back within an hour closely resembled work her Google team had been iterating on for nearly a year. The admission did not come from an outsider or a casual observer, but from a senior engineer at one of the world’s most influential technology companies.
A problem Google had been Fighting with
The task Dogan gave Claude Code was far from trivial. It involved designing distributed agent orchestrators, a class of systems responsible for coordinating multiple AI agents so they can work together toward a shared goal. These orchestrators are considered a foundational piece for next-generation AI systems, especially as models move beyond single-task responses and toward collaborative, multi-agent workflows.According to Dogan, Google has explored several architectural approaches to this problem since last year. Like many large organizations, the company faced internal debates, competing ideas, and a lack of consensus on what the final design should look like. The work was ongoing and unresolved.
That context is what made the result from Claude Code so striking.
“I’m not joking and this isn’t funny,” Dogan wrote candidly. “We have been trying to build distributed agent orchestrators at Google since last year. There are various options, not everyone is aligned… I gave Claude Code a description of the problem, it generated what we built last year in an hour.”
 |
| Source: X.com |
Her words resonated because they captured a quiet anxiety—and excitement—felt across the industry: that AI systems are beginning to compress months or years of human effort into astonishingly short timeframes.
A simple prompt, surprising depth
One detail that stood out to many readers was how Dogan framed the experiment. She did not feed Claude Code proprietary designs or internal Google documentation. In fact, she went out of her way to avoid doing so.Because she could not use confidential information, Dogan created a simplified version of the problem using publicly known concepts and existing ideas. The prompt itself was modest—just three paragraphs outlining the challenge. There were no step-by-step instructions or detailed architectural constraints.
Even with those limitations, Claude Code produced an output that mirrored much of what her team had arrived at through months of discussion, prototyping, and revision.
Dogan was careful to add nuance. The result, she said, was not flawless. It still required refinement and would not be dropped directly into production. But the core structure and reasoning were close enough to raise serious questions about how quickly AI coding tools are maturing.
A message to skeptics
Rather than framing the experience as a threat, Dogan offered it as a challenge to skeptics of AI-assisted programming. She encouraged developers who doubt the usefulness of coding agents to test them not on toy problems, but in domains where they already have deep expertise.That advice reflects an important point: the value of these tools becomes clearer when experienced engineers can critically evaluate their output. AI may not replace human judgment, but it is increasingly capable of producing strong first drafts—sometimes at a level that surprises even experts.
Competition without hostility
The conversation around Dogan’s post quickly turned to questions about competition between major AI labs. When asked whether Google uses Claude Code internally, Dogan clarified that it is permitted only for open-source projects, not for internal development. This distinction reflects ongoing concerns about data governance, intellectual property, and security within large organizations.Another user asked the inevitable question: when would Google’s own models, particularly Gemini, reach a similar level of performance? Dogan’s reply was brief but telling. “We are working hard right now,” she wrote. “The models and the harness.”
Importantly, she rejected the idea that AI development is a zero-sum game. Recognizing strong work from competitors, she argued, benefits the entire field. “Claude Code is impressive work,” she added. “I’m excited and more motivated to push us all forward.”
That perspective aligns with a growing consensus among senior researchers: rapid progress in AI often comes from shared ideas, healthy competition, and mutual acknowledgment rather than secrecy alone.
How fast AI coding has evolved
Dogan also used her post to reflect on the broader trajectory of AI-assisted programming—a timeline that underscores why this moment feels so disruptive.2022:AI tools could reliably complete individual lines of code.
2023: They began handling entire functions or sections within a file.
2024: Systems became capable of working across multiple files and assembling simple applications.
2025: Today’s tools can generate, refactor, and restructure entire codebases.
This progression, she noted, outpaced even her own expectations. In 2022, she did not believe the 2024 milestone could realistically scale into a global developer product. A year later, she still thought current capabilities were at least five years away.
“Quality and efficiency gains in this domain are beyond what anyone could have imagined so far,” Dogan wrote.
Signals from the Viral Moment
The post quickly went viral, drawing more than four million views and thousands of reactions. Many comments focused less on Claude Code itself and more on what the story symbolizes.One user observed that AI tools may be sidestepping the bureaucratic inertia that often slows innovation in large organizations, potentially enabling a new wave of individual creativity. Another predicted that code generation and feature development would soon accelerate at a pace developers are still unprepared for.
These reactions highlight a shifting narrative. AI is no longer seen only as an assistant that saves time on repetitive tasks. Increasingly, it is viewed as a collaborator capable of proposing viable system-level designs—something once thought to require prolonged human deliberation.
Why This Moment Matters for Software Developers
For engineers, Dogan’s experience offers both reassurance and a wake-up call. AI tools are not perfect, and they still depend on human expertise to guide, validate, and improve their output. At the same time, ignoring their capabilities may soon become a competitive disadvantage.
The real lesson may not be that AI can replace a year of work in an hour, but that it can dramatically shorten the path from idea to implementation. Teams that learn how to integrate these tools thoughtfully—while maintaining rigorous standards—are likely to move faster and explore more ambitious ideas.
As Dogan’s post makes clear, even those building the future of AI are still being surprised by how fast that future is arriving.
