There's something almost surreal about watching Hiroki Tomiyasu review software architecture while steering a tractor across frozen Hokkaido earth. One hand on the wheel, the other scrolling through lines of code on his phone—code he didn't write himself but orchestrated into existence using plain Japanese and an AI tool called Codex.
Tomiyasu isn't what you'd call a typical early adopter. He's a former civil servant turned farmer, managing 100 hectares of broccoli, pumpkins, green onions, and soybeans across Biratori and Mukawa. No programming background. No computer science degree. Just a growing impatience with expensive agricultural software and the stubborn belief that there had to be a cheaper way.
Turns out, there was.
The ¥60,000 Alternative
When Tomiyasu started pricing out commercial greenhouse automation systems—the kind that remotely control ventilation roll-ups across multiple structures—quotes came back around ¥1,000,000 (roughly $6,300). For a farmer working at scale but without venture capital backing, that's a non-starter.
So he spent about ¥60,000 to ¥70,000 ($375–$440) on hardware components instead: an ESP32 microcontroller, a BTS7960 motor driver, some sensors. Then he fed instructions to OpenAI's Codex, which generated not just the code but the wiring diagrams. Two months later, he had a working system.
The economics are striking. A 15-to-1 cost ratio doesn't just make automation more accessible—it fundamentally changes the calculation of when it makes sense to automate at all.
Business Insider Japan visited the farm and documented the setup in detail, with cooperation credit to OpenAI Japan: the Cloudflare Workers backend, the LINE bot interface that became his operational command center, the temperature sensors scattered across greenhouses. What the reporters found wasn't a prototype or proof-of-concept. It was production infrastructure, stable enough to run during a media tour.
Building While Plowing
The LINE bot evolved into something of a digital farmhand. It checks individual greenhouse temperatures using SwitchBot sensors. Opens and closes vents remotely. Pulls up work schedules from the farm database. For a crew that includes workers speaking different languages, Tomiyasu built a translation bot—also via Codex—that handles real-time communication in group chats.
Then there's the GPS work logging app, which he claims to have built "in about a day" from the tractor seat. Workers and equipment now automatically log time and field locations via smartphone GPS, creating digital records without manual timesheets or paper trails.
That last detail strains credulity a bit, perhaps. Building functional software in a day while also plowing fields suggests either remarkable prompting skill or some narrative compression. But even if it took a week, that's still a dramatically compressed development cycle compared to traditional procurement timelines.
The Codex Evolution
Tomiyasu's workflow became viable because of specific product developments in OpenAI's Codex offering. The tool evolved from a developer-focused API into a mobile-accessible preview feature that lets users review and redirect AI-generated work from their phones while the heavy lifting happens on a connected Mac or remote host. That capability—intervening "when it matters" from wherever you happen to be—is what makes tractor-seat development theoretically possible.
OpenAI has been public about usage growth. As of June 2, 2026, the company reported more than 5 million weekly users, and rolled out role-specific plugins positioned as "no coding required" workflows across various professional domains. The ChatGPT Pro Community, an OpenAI-affiliated publication, featured Tomiyasu's work with the actual prompts he used, component lists, and architectural decisions laid bare.
Which raises a fair question about how organic this adoption story really is versus how showcased it might be. The Business Insider Japan piece includes cooperation credit for OpenAI Japan. That doesn't invalidate the technology or the implementations, but it does suggest these examples are being surfaced and amplified as part of a broader positioning effort.
Beyond Greenhouse Vents
Tomiyasu's automation ambitions extended into decision support as well. He photographs broccoli plants showing potential disease symptoms and asks ChatGPT for likely diagnoses and triage guidance. He pulls NDVI vegetation indices from satellite data and overlays them on custom field maps to monitor crop health. Before investing in RTK-GPS auto-steer hardware for tractors—a significant capital expense—he used prompts to understand the technology well enough to make an informed purchasing decision.
Even mundane documentation got the AI treatment. GPT's image annotation auto-labels control panel photos in Japanese. He's experimented with 3D-printing custom parts by having Codex generate designs.
What stands out is the technical diversity: hardware integration, API orchestration, database management, user interface design. All stitched together by someone describing outcomes rather than writing functions. It's a fundamentally different relationship to software.
The Broader Bet
The proposition being tested here is straightforward, if ambitious: Can AI agents handle the entire technical stack for operational tools if the user can clearly articulate the business logic?
Tomiyasu's greenhouses don't care whether the ventilation controller was written by a professional developer or generated through prompts. The LINE translation bot doesn't perform worse because its creator can't code. The GPS logging system captures the same data it would if built by a consultancy—maybe better, because the person designing it intimately understands the operational context.
The competitive landscape reflects similar bets. Anthropic launched Claude Code with desktop and app workflows, though it's faced recent security questions in some markets. Replit positioned its Agent 4 product around building apps from natural language with "no code needed." GitHub Copilot, long a developer tool, has been adding agent-style workflows through Workspace and CLI updates.
All these products are chasing the same insight: there's a vast gap between people who can describe what they need automated and people who can technically implement that automation. Close that gap, and you unlock operational efficiency at scales previously uneconomical.
What He Didn't Automate
It's worth noting the limits. Tomiyasu didn't build mission-critical systems where failure could endanger crops or livestock. His tools augment existing processes—greenhouse vents can still be opened manually; GPS logs supplement rather than replace human oversight. The automation is incremental, not existential.
There's also the maintenance question. Tomiyasu appears to iterate continuously, refining prompts and adjusting integrations as conditions change. That works when the builder is the primary user and has intimate knowledge of the operations. It's less obvious how these systems would transfer to someone else or scale across multiple farms without that direct involvement.
You could argue that's actually a feature, not a bug. Owning your automation stack means controlling your data and minimizing recurring costs after initial deployment. It's a different ownership model than subscribing to vertical-specific SaaS platforms—though it requires a different kind of ongoing engagement.
The Hokkaido Context
Hokkaido agriculture operates at larger scales than most Japanese farming regions. Research from NARO Hokkaido notes that 100-hectare family farms have emerged in the region, meaning Tomiyasu's operation isn't a local outlier even if it sounds large by broader standards.
What is unusual is the degree of custom automation, typically associated with industrial operations or well-funded agtech startups. If a single non-technical operator can replicate that capability using AI agents and commodity hardware, the financial threshold for when automation makes sense drops considerably.
Tomiyasu's toolkit—greenhouse control, multilingual communication, GPS tracking, satellite monitoring, disease diagnosis—reads like a complete SaaS stack. Except he built it, he controls it, and the recurring costs are minimal. That's a fundamentally different economic model.
The Unanswered Question
The broader implication here isn't really about farming. It's about whether operational leaders in any domain can now bypass procurement processes, implementation partners, and multi-month deployments in favor of describing what they need and letting an AI agent handle technical execution.
Tomiyasu's 100 hectares are one data point. The more interesting question is how many others are running similar experiments in logistics, manufacturing, retail, or healthcare—and what the aggregate effect looks like when non-technical operators become their own automation engineers.
Maybe there's a whole parallel ecosystem emerging, built by domain experts using AI tools, running quietly beneath the radar of traditional enterprise software vendors. Or maybe these are still edge cases, showcased examples that don't yet represent widespread adoption.
Either way, the fact that it's even possible for a farmer to build enterprise-grade automation from a tractor seat suggests something fundamental has shifted. Whether that shift is permanent—and how far it extends beyond early adopters—remains very much an open question.
