# The Future of Software Development: Embracing AI and Quality > Software engineering veteran and Gas City CEO Chris Sells on AI coding agents, redefining software quality, and building "software factories," on Very Good Engineering. - Source: https://verygood.ventures/blog/chris-sells-gas-city-future-of-software-development-ai-and-quality/ - Published: 2026-07-08 - Author: VGV Team - Tags: Flutter, AI, Software Engineering, Very Good Engineering --- {% video url="https://www.youtube.com/watch?v=FWh9TCSFdZw" title="Chris Sells: The Future of Software Development, Embracing AI and Quality" %} Few people have watched software development change shape as many times as Chris Sells. In this episode of *Very Good Engineering*, host Jorge Coca sits down with the industry veteran, prolific author, and current CEO of Gas City to talk about what actually changes when AI agents start writing the code, and what doesn't: the need for genuinely high-quality software. ## About Chris Sells Chris has spent decades at the center of the industry, with stops at Intel, Microsoft, Google, Meta, and Sourcegraph. At Google he spent three years on the Flutter team, helping take it from version 1.0 to 3.0 and from hundreds of thousands of users to millions. At Meta he worked on developer experience for AR and VR in Reality Labs. At Sourcegraph he helped ship version 1.0 of the AI coding assistant Cody, for both consumer and enterprise. Along the way he wrote more than a dozen books on software development. He got his start decades ago writing database software on an Apple II Plus in his mother's basement, as a teenager keeping track of his Dungeons & Dragons characters. Today he is CEO of Gas City, an open-source platform for orchestrating AI coding agents into software factories. ## From Typing Every Line to Directing Agents Jorge opens by asking what feels genuinely new about AI, versus what's just an old problem in a new costume. Chris's answer is direct: the problem has never changed. "We have software-shaped problems, and we need high-quality software solutions," he says, a line he returns to throughout the conversation. What's changed is the method. For his entire career, no matter how sophisticated the tools around it got, compilers, debuggers, profilers, autocomplete, someone typed every line of every file by hand. AI coding agents broke that. For the first time, he says, humans are neither writing the code nor, at the edges of agentic engineering, even reading it. That shift produced what the industry now calls vibe coding: describe what you want, get code that runs. Chris draws a hard line between that and what real engineering requires. Vibe coding gets you something that works on your machine. Production software is a feature or fix you can check into an existing enterprise system with confidence it holds up under load, meets its latency requirements, produces the side effects you want and none you don't, and integrates cleanly with your logging and escalation systems. "Building real enterprise software requires way more than 'it works on my machine,'" he says. "That is the core difference." ## What "Quality" Actually Means Jorge presses on a gap he sees across the industry: everyone talks about quality, but nobody agrees on what it is. Chris agrees there's no shared contract for it, even after decades of trying. Engineers have built up plenty of principles: don't repeat yourself, single responsibility, loose coupling and high cohesion, cyclomatic complexity as a way to measure how tangled a piece of code has become. But applying them has always depended on a human reading code and sensing when something is off, what engineers call a "code smell." Agents can't smell anything. If quality lives only in engineers' heads, it can't be handed to an agent. That is why Chris thinks writing those standards down, explicitly and specifically, matters more now than ever. Skip a principle like DRY and the consequences are the same as they've always been, just faster: an agent fixes a bug in the one place it happened to look, and leaves the same bug sitting in every other place the logic was copied. "You still want that same code quality, those same things that we have figured out as software engineers that bring high-quality software," he says. "We want all of that in our code." ## Inside Gas City: Building a Software Factory Chris's answer to the standards problem is to encode them directly into the system that runs the agents. Gas City, built on ideas from Steve Yegge's Beads and Gas Town projects, gives engineers a small set of primitives for that: Beads track units of work, formulas define repeatable multi-agent workflows (say, a code-review loop or an architecture-compliance check), and a rig is a project that's been fully configured with agents, skills, and those formulas, turning it into its own miniature factory. The numbers give a sense of scale. Gas City itself runs on Gas City: 23 public repositories, roughly a hundred agents running around the clock, more than a hundred community issues and pull requests in flight at once, and a peak of over 70 pull requests merged in a single day. Chris calls the shift a version of the industrial revolution: moving from handcrafting software, to handcrafting the prompts that generate software, to automating the prompts themselves. One trick he's found valuable is running the same quality check through many different agents at once. Because each one is trained differently, each has its own point of view. Run two to thirty of them over the same code and the genuine issues show up again and again, while one-off hallucinations cancel out as noise. ## Where Humans Fit: Setting Intent, Not Syntax If agents are doing the typing, what's left for the engineer? Chris's own workflow is telling. When a new issue or pull request comes in, he spends his time on one question: does this belong, does it serve the project's actual intent? He deliberately does not spend that time judging architecture, implementation quality, or test coverage. Once he decides something belongs, he applies a label and walks away. From there, formulas take over: verifying architecture, writing and running tests, checking against the standards the team has already encoded, until the change clears the bar and can be merged. The consequence, he argues, is a real change in what a software engineer's job looks like. Less time goes into implementing individual features by hand, and more into building and operating the factory itself: the formulas, the quality gates, staging, releases, and watching production logs for the kind of drift that signals something needs attention. It's a bigger job than it sounds, since a real software factory has to cover the whole lifecycle, not just the moment code gets written. ## Agentic Teammates, Beyond Code Gas City isn't only for shipping code. Chris and his small team also use it to help run the company itself, through a set of named agentic teammates with their own skills, responsibilities, and memory: Seth handles site reliability, Randy manages releases, Maya works marketing, Priya handles user experience. They're reachable through Discord, Slack, SMS, and email, and humans and agents can sit in the same conversation together. Want to move the website forward? Bring in Maya, Priya, and the team's legal advisor at once, and get three different, occasionally conflicting, points of view before deciding what to do. He's also candid that agentic engineering today is largely a single-player sport: one person, one terminal, one set of agents. Making it multiplayer, so a whole team works from the same factory, the same formulas, and the same standards, is one of the features he and his team are actively building next. ## Taste, Judgment, and the Human Agenda The question Chris says he hears most from working engineers is blunt: if AI writes the code, what does the industry need me for? His answer centers on taste and agenda. Agents, he says, will do whatever they're asked, to the best of their ability, but they don't want anything. They have no stake in the outcome and no way to judge whether "done" is actually good. That judgment, he argues, still belongs to software engineers, UX designers, and product people, the ones who know what a great outcome looks like in their own discipline. Agents don't replace that judgment; they let the people who have it move much faster. It's a philosophy that lines up closely with one already familiar at VGV. Jorge notes that VGV CEO David DeRemer frames the same idea as trust, taste, and tempo, and Chris's take lands in the same place from a different direction: the tools accelerate execution, but a human still has to define what "good" means and hold the line on it. ## What's Next Asked to name a Flutter package he wishes he'd built, Chris doesn't hesitate: something for AI-generated, dynamic UI. He's experimented with it and thinks Flutter is the ideal platform for it, since it can run anywhere the interface needs to live. He's blunt about where the technology stands today, though: when asked for one word to describe the current state of AI-generated UI, he picks "slow." He also sees far more potential in tailored, AI-driven interfaces than in the text-box chat experience most products default to today. For his own work, Chris turns to Claude for planning and design conversations and to Codex for the implementation that follows, across front end, back end, and databases alike. As for when software factories go from novelty to norm, his prediction is soon: he expects this to be the year they're introduced widely, and the following year the one where they become the dominant way software gets built. ## Key Takeaways - **The goal hasn't changed, only the method.** Software-shaped problems still need high-quality solutions. What changed is that agents, not humans, now write and often even read the code. - **Vibe coding and production software are not the same thing.** "It works on my machine" is a much lower bar than an enterprise system needs to clear. - **Quality has to be written down, not just felt.** Agents can't sense a code smell, so principles like DRY and SOLID need to become explicit, encoded standards. - **Humans should focus on intent, not implementation.** Decide what belongs in the system; let agents and their quality gates handle the how. - **Taste and agenda are what agents don't have.** They'll execute anything you ask; deciding what's worth asking for is still a human job. ## Conclusion Chris Sells has spent 40 years watching the tools of software development change shape, from punch cards to text files to, now, agents that write the text files themselves. His take is refreshingly grounded: the tools change, but the discipline doesn't. Until AI becomes something more than a tool, and he's careful to say we're not there yet, the job of deciding what's worth building, and holding it to a real standard of quality, still belongs to the people doing the building.