When Your AI Agent Ships Code While You Sleep

Ashe Magalhaes runs a Slack workspace where autonomous agents manage her entire product development pipeline. Each channel represents a different agentic workflow. Error notifications arrive throughout the day. When aesthetic.video—a collaborative video editing platform she built—started showing an infinite load bug on the drafts page, she typed a single instruction: "Kick off a task to fix the infinite load on the drafts page." Then she moved on.

"I trust 5.4 will figure that out," she told OpenAI's Romain Huet during a recent interview.

This is what software development looks like when you delegate entire features to AI agents that can actually ship production code. Not prototypes. Not demos. Actual products that Magalhaes uses daily and that other people pay for.

The Hearth AI founder—who previously built machine learning systems at Airbnb and worked on satellites at NASA—represents an emerging category of solo builders who treat AI models less like assistants and more like engineering teams. Her workflow raises questions about what "building" means when the builder's primary skill becomes directing autonomous systems rather than writing code.

The Speed Problem

Magalhaes built Hearth AI in 2022 as "the first agentic CRM," attempting to create AI workflows on GPT-3.5 when, as she puts it, "the underlying systems were still pretty fragile." The technical challenge wasn't just making non-deterministic systems work—it was explaining to users and investors why the team kept rebuilding their entire stack every few weeks.

"The history of tech had been, okay in ML you maybe spend a year laying the data pipes, and then and only then can you do something predictive and interesting with the ML product," she said. "And for us, it was like, no we want to surface an experience to users that is flexible."

That flexibility came with chaos. Consumer behavior around querying agents didn't exist yet. Users didn't understand that Slack bots could respond to messages—they thought bots only sent outbound notifications. The engineering patterns were being invented in real-time by a loose network of builders sharing tricks.

What's changed in two years isn't just model capability. It's reliability. Magalhaes can now tell OpenAI's Codex to extract a prototype from her personal website and spin it into a standalone product—and it works. In one shot. Six weeks ago, that wasn't possible.

"I think that's kind of a huge difference, even going back like six weeks ago," she noted.

The Secrets Page

Magalhaes maintains a public-facing website at ashe.ai that showcases her portfolio and products. Behind it sits a "secrets page" where she prototypes ideas before they face users.

The workflow: She builds features within her personal system using Codex. If she actually uses the feature daily, she posts it on X (formerly Twitter). If people react, she instructs the AI agent to "break off this chunk" into either a standalone product or an open-source repository.

This is building in public, but with a crucial filter. The public only sees what survives her own daily use. Everything else stays in the lab.

Take aesthetic.video, the collaborative editing platform. It started as a prototype on her secrets page because she needed a way to share interview footage with guests while letting them comment on specific timestamps. She tested it herself. Posted about it. Got requests for access. Then told Codex to extract it as a standalone product.

The entire platform now includes Google storage for videos, Gemini embeddings for semantic search ("search for people laughing" actually works), automatic X integration, and an agentic clipper for creating social media reels. She built it while solo.

"The name of the game is speed," Magalhaes said. "The key thing is you could build so much software nowadays, but what will you build that you yourself use every day that you find to be a beautiful and elegant experience?"

That last qualifier—beautiful and elegant—matters. She's deliberate about distinguishing her work from what she calls "vibe coded products" that break under real use. Speed without reliability is just technical debt at scale.

The Definition Problem

Magalhaes's background reveals the tension in current conversations about AI and engineering work. She drove a solar-powered car she helped build across the Australian Outback at 90 kilometers per hour. Teams were crashing. Cambridge's entry crashed. She spent two years building something she would literally put her body into.

"I was very aware that I had spent two years building something with a team that I would put my body into and that level of rigor, like really played into later experiences," she said.

That level of rigor—understanding systems deeply enough to trust them with your physical safety—represents one definition of engineering. But Magalhaes also describes herself using different language: "I view the builder archetype as quite close to the artist archetype. I think I'm happiest viewing myself as like a conduit for creativity."

When builders become artists and artists become builders, when the primary skill becomes taste and direction rather than implementation, what exactly are we measuring?

The transcript cuts off mid-sentence, but Huet's unfinished question captures the moment: "The role of being an engineer is changing. Like, what's your view on all of t—"

We don't get Magalhaes's answer, but her workflow suggests one. If agents can reliably ship production code, engineering becomes less about implementation and more about judgment. Which features matter? Does this solve a real problem? Would I actually use this daily? Is it beautiful?

Those questions require human discernment that currently can't be automated. But they're also questions that don't require deep knowledge of algorithms or system architecture. They require taste, psychology, and what Magalhaes calls "relational intelligence"—understanding how people actually connect with technology.

The Regulatory Vacuum

From a policy perspective, this development pattern exists in an interesting void. There's no regulatory framework that contemplates autonomous agents shipping production code. Software liability, professional licensing, quality assurance standards—all assume humans are writing the code and making the technical decisions.

When an AI agent introduces a bug that causes data loss or security breaches, who's liable? The person who gave the instruction? The company that made the model? The model itself?

These aren't hypothetical questions. Magalhaes is already delegating critical infrastructure decisions to agents. When her aesthetic.video platform threw errors, she didn't debug them herself—she instructed an agent to fix them and moved on.

This works until it doesn't. And when it doesn't, we'll discover whether our existing legal and regulatory frameworks can handle software development where the developer's primary contribution is writing English instructions rather than code.

The speed advantage is real. Magalhaes mentions building an Oura ring integration, posting her wake-up times publicly, and having someone suggest a social accountability version. "The magic of this is now I can quickly sprint at this without it taking up so much time," she said.

Sprint is the operative word. We're optimizing for velocity in an environment where we haven't yet figured out the safety rails. That's not necessarily wrong—innovation requires some tolerance for risk—but it's worth naming explicitly.

What Gets Built

Magalhaes's thesis about "relational intelligence" cuts against the current moment in an interesting way. She believes "AI should augment the human experience and allow us to feel more connected in the time that we have." Her technical work aims to extend human capacity for maintaining relationships.

But her workflow demonstrates the opposite dynamic. She builds alone, in her Slack workspace, instructing agents. The collaboration happens with the AI, not with other engineers. The building in public she describes is really building in private and then publishing results.

There's no contradiction here—she's clear-eyed about the tradeoff. Years of building "feeling quite disconnected" taught her something. Now she's trying to build tools that create connection while using tools that enable solo development.

Whether that's sustainable—whether one person directing AI agents can compete with larger teams in the long run—remains an open question. The economics suggest it should work. The historical pattern of software development suggests coordination costs might still matter.

What's certain is that Magalhaes represents a real phenomenon: capable solo builders using AI agents to ship production software at speeds that would have required entire teams 24 months ago. How widely that pattern scales, and what it means for how software gets built, will shape both the technical and regulatory landscape over the next few years.

Samira Okonkwo-Barnes is Buzzrag's Tech Policy & Regulation Correspondent.