Building Frontier Products with AI at One-Person Scale
AI-native product development · 2026 · 5 min read
A field report on directing agents across product, design, engineering, infrastructure, and QA while keeping architecture, verification, and release decisions human.
AI proficiency is not the ability to produce a large volume of code. It is the ability to direct software agents across the boundaries of a complete product, make the decisions they cannot own, and know what evidence is strong enough to ship.
I am testing that proposition by building at one-person scale across consumer software, native applications, streaming media, publishing infrastructure, private agentic systems, emulation, and spatial computing. AI participates from research through operations. I remain the single human responsible for the product model, architecture, integration, verification, release, and result.
The products are evidence of the production model. They are deliberately different because repetition inside one familiar stack would prove less.
Six systems, six different boundaries
| Product | Frontier system | Where agents participate | Human verification boundary | | --- | --- | --- | --- | | PartyParty | A native macOS capture path packages system audio as AAC over LL-HLS, then selects direct venue Wi-Fi or a cloud relay while guests listen through native AVPlayer on locked iPhones. | Apple platform research, Swift, Go streaming services, browser code, Cloudflare infrastructure, test design, and documentation. | Fresh Macs and iPhones, AirPods, background and lock-screen playback, direct and relayed paths, and real venue networks. | | Darabiha | A private archivist reads multilingual evidence, reconciles people and relationships against a typed graph, and proposes audited operations instead of free-form answers. | Legacy archive reconstruction, graph modeling, privacy UX, typed tools, full-stack implementation, migration, and testing. | Family members remain the source authority. Ambiguity and contradiction fail closed, and the records remain private. | | Retro Porting Toolkit | A publishing graph connects static recompilers, runtimes, native ports, hardware targets, games, releases, and contributors in an agent-readable content system. | Information modeling, source normalization, application and CMS implementation, media processing, search, and verification. | Primary project sources govern attribution and status. I built the website and publishing system, not the recompilers or ports it documents. | | Chiptunes.app | One symbolic score drives a procedural composer, register-level DMG audio emulation, reactive games, an editor, a permanent stream, and a compiler that emits a bootable 32 KB cartridge. | Composition systems, emulator and compiler implementation, games and shaders, parity harnesses, streaming, and deployment. | Musical judgment stays human. Spectral checks compare browser playback with the cartridge running inside a CPU emulator. | | ramine.net | A typed content graph compiles into two identity-isolated publications, fully prerendered routes, durable citations, optimized media, and client-side full-text search. | Research synthesis, schema and interaction design, application and WebXR engineering, media work, testing, deployment, and publishing operations. | Claims, citations, identity boundaries, accessibility, browser behavior, build isolation, and every release remain reviewable and testable. | | Spatial Gallery | One React Three Fiber scene runs as a gaze-and-pinch room-scale WebXR gallery on Apple Vision Pro and as a mouse and touch experience in ordinary browsers. | Spatial interaction prototypes, WebXR input and locomotion, rendering, video delivery, performance work, browser fallbacks, and regression tests. | Comfort, native input, visual correctness, and the 90 fps headset budget require physical Vision Pro and browser testing. |
The operating system around the agents
The leverage comes from making the work legible enough that agents can contribute without being allowed to redefine the product accidentally.
The product contract comes first. Each repository records its user problem, system boundaries, platform constraints, ownership, and release conditions. Agents receive bounded work against that contract instead of an open invitation to generate an application.
Architecture is expressed as enforceable structure. Typed records, validation, fixtures, tests, content eligibility, explicit site selection, and agent-readable instructions turn decisions into constraints the next task must respect. A handoff records facts about the live system and its remaining risks so a different agent can resume without inventing a new baseline.
Implementation and review are separate jobs. Agents can research unfamiliar APIs, implement across languages, generate tests, inspect failures, and review one another's work. The output is still judged against the product behavior and the underlying platform, not against whether the code looks plausible.
Verification follows the risk. A type check is useful evidence for a schema change. It is not evidence that Safari drew the correct native cursor, an iPhone continued playing after its screen locked, a cartridge sounds like the browser renderer, or a spatial interaction remains comfortable in a headset. Those boundaries are tested on the real system.
Operation is part of the product. Deployment, media processing, metadata, source attribution, privacy, monitoring, and continued iteration are not cleanup after the demo. They are part of the work agents are expected to support and the human owner is expected to verify.
Where the method fails if used carelessly
Agents are persuasive before they are correct. They can preserve a stale feature claim after the feature has been removed, accept computed browser state as proof of visible native behavior, or make a collaborative ecosystem sound like one person's technical work. All three failures produce polished but false stories.
The countermeasure is not to remove AI from the process. It is to bind every claim to the live system, separate contribution from attribution, keep irreversible decisions human, and require evidence at the layer where the user experiences the product.
That is the production model I am exploring: one experienced product builder using agents to own a much larger system, without lowering the standard for usefulness, performance, reliability, or finish.