Patrick Ma Built AI Agents Before the Industry Knew What to Call Them

The fear that AI will replace software engineers has become one of the loudest debates in technology. New tools can generate code, fix bugs, explain unfamiliar files, and build simple applications with a few lines of natural language. To some people, that sounds like the beginning of the end for programmers. To Patrick Ma, senior AI engineer and Quora’s first dedicated AI agent engineer on Poe, it sounds like a misunderstanding of what good engineers actually do. “AI can write code,” Ma says. “But it cannot take responsibility for the system.”

That distinction has shaped much of Ma’s recent work. At Quora, he architected the first AI agent on Poe, the company’s AI product, before the software industry had settled on a shared meaning for the word “agent.” He has worked on Poe for more than three years, moving from iOS engineering to agentic AI work as the product and the market evolved together. “I built agents before ‘agent’ had a real definition,” Ma says. “That meant we could not rely on buzzwords. We had to focus on what the product needed to do for users.”

Poe gives users access to leading AI models and lets them create AI-powered apps and experiences. Ma served as tech lead on major Poe launches covered by TechCrunch, including Previews, App Creator, and group chats across models. Those products reflect a broader shift in software creation: more people can now describe what they want a computer to build, even if they do not write traditional code themselves.

That is one reason Ma does not believe the future is simply fewer engineers. He believes the cost of building software is falling, which may increase demand for useful software. The easier it becomes to start building, the more the industry needs engineers who can make systems reliable, maintainable, and worth using. “Basic software tasks are much easier now,” Ma says. “But complex software still needs architecture. It still needs taste. It still needs someone who understands what should be built and how it would hold together.”

This is where Ma’s view becomes more precise than the common argument about AI replacing jobs. AI can generate large amounts of code, but code volume is not the same as good engineering. A system has to be reviewed, maintained, secured, evaluated, and understood. It has to fit into a codebase other people can work with. It has to behave correctly in critical situations.

“The best engineers will not be the ones who avoid AI,” Ma says. “They will be the ones who use it aggressively without surrendering judgment.”

Ma speaks from unusually direct experience. He currently writes more than 99 percent of his code using AI. That does not mean he has stopped engineering. It means his role has shifted toward directing, reviewing, structuring, and deciding.

“When I say I write code with AI, I do not mean I stop thinking,” he says. “I think more about the shape of the system, the constraints, and whether the output is actually correct.”

Inside Quora, Ma became a champion for AI coding tool adoption. In April 2025, he helped drive a move from Cursor to Claude Code at a time when he says much of the industry had not yet recognized the coding agent wave. His advocacy helped increase Claude Code usage by 20 times in two months.

That work was not just about telling engineers to use a new tool. Ma also worked on making engineering itself more AI-native. He restructured parts of the codebase and even helped redesign the engineering interview process for an environment where AI tools were becoming part of normal development. He also drove organizational adoption of agent evaluation tooling.

“You cannot just give people AI tools and expect productivity to appear,” Ma says. “The codebase, the process, the evaluation, and the culture all have to change.”

That is one of the lessons he thinks the industry is still learning. AI adoption can become shallow when companies only measure usage. Ma has seen how token consumption or leaderboard-style incentives can backfire. People may optimize for the metric instead of the work, even setting up meaningless automated prompts to increase their AI usage numbers.

“AI usage is not the same as AI productivity,” he says. “If you measure the wrong thing, people will optimize for the wrong thing.”

The same caution applies to AI-generated code in serious systems. Ma believes engineers should use AI, but he also believes they must remain responsible for what reaches production. He has seen cases where AI pushed bad code on behalf of engineers without enough human awareness. In critical systems, that cannot become acceptable.

“AI agents are still early,” Ma says. “They are not reliable all the time. In real-world systems, humans still have to review and take responsibility.”

That responsibility may become more important as software creation becomes easier. Poe’s coding agents are part of a larger movement toward letting users create software through natural language, but Ma believes that accessibility does not erase the need for engineering discipline.

A non-engineer may be able to generate a working app. An engineer still has to understand whether the code is safe, scalable, maintainable, and aligned with the product’s real purpose. The difference between a quick output and durable software still matters.

“Software is not only about whether something runs once,” Ma says. “It is about whether it keeps working, whether people can change it, and whether the system can survive real use.”

That belief also shapes how Ma evaluates the broader AI market. His work as a venture scout and hackathon judge has sharpened the way he separates useful AI projects from impressive ones. The question, for him, is not whether a demo produces surprise. It is whether the idea can become a system people trust.

“Whether you are judging code or judging founders, you are asking a similar question,” Ma says. “Is this real? Is this useful?”

His next role will bring that question even closer to the center of AI software development. Ma is joining Cognition, maker of Devin, the first AI software engineer. At Cognition, his next chapter stays inside the same technical question: how much of software engineering can AI take on, and what must still be governed by human judgment?

For Ma, the future of software engineering is a higher standard for using it well. Engineers may write fewer lines by hand, but they will need stronger judgment about architecture, product value, safety, and maintainability.

“AI lowers the cost of building,” Ma says. “It also raises the bar for what engineers are responsible for.”