個人1名+AIエージェント2種で金融グレードシステムを全自動運用している話A developer shares how they operate a production-grade financial system solo by…

A developer has published a detailed account of running a production-grade financial system single-handedly, relying on two AI coding agents to write and propose most of the changes. The write-up, posted on the Japanese engineering platform Zenn, is notable because it describes not just AI generating code, but AI agents operating inside a live production workflow where they open pull requests that a single human reviews and merges.

The system, referred to as fin-os-prod, runs on Cloudflare Workers, the edge serverless platform that executes code close to users rather than in a traditional centralized server. The excerpt indicates the stack pairs Workers with what is likely Cloudflare D1, the platform's SQLite-based database, though the author's full architecture description would confirm the exact components. Building a "financial-grade" system on this foundation suggests an emphasis on reliability, correctness, and auditability, qualities that matter when software handles money or financial data.

At the center of the setup is a multi-agent arrangement combining GitHub Copilot coding agent and Claude coding agent. These are distinct from the inline autocomplete that first made Copilot famous. A coding agent can take an assigned task or issue, work through a codebase, and produce a complete set of changes as a pull request, effectively acting as an autonomous contributor rather than a typing aid. GitHub's coding agent, introduced as an evolution of Copilot, is designed to pick up issues and return draft PRs. Claude's coding capabilities, delivered through Anthropic's models and tooling, similarly can plan and edit across multiple files.

The developer's workflow appears to place the human firmly in the role of reviewer and gatekeeper. According to the account, the agents automatically open pull requests, and the single human operator reviews and merges them. This human-in-the-loop pattern is the crucial safety mechanism: rather than allowing AI to deploy changes directly, every modification passes through the same code review and merge gate that a conventional engineering team would use. For a system described as financial-grade, that checkpoint is likely essential to catching errors before they reach production.

Using two agents rather than one is the other distinguishing choice. The author frames this as a multi-agent development setup, and while the excerpt does not fully detail the division of labor, running Copilot's agent and Claude's agent in parallel could allow the developer to compare approaches, assign different tasks to each, or use one agent's output as a check on the other. It also reflects a broader industry pattern in which teams increasingly mix models and tools from different vendors rather than committing to a single provider.

The account sits within a fast-moving shift in how software gets built. Over the past year, several vendors have pushed coding assistants from suggestion engines toward more autonomous agents. GitHub, Anthropic, and others have released agent modes capable of executing multi-step work, and platforms such as Cloudflare have leaned into serverless primitives that pair naturally with automated deployment pipelines. What remains comparatively rare, as the author notes, is a documented case of these agents running the development loop for a production financial system rather than a demo or side project.

Several caveats are worth keeping in mind. This is a single practitioner's report rather than an independently verified case study, and terms like "financial-grade" and "production" are the author's own characterizations. The long-term maintainability, security posture, and regulatory considerations of a solo-operated financial system driven largely by AI-generated pull requests are not addressed by the excerpt and would require deeper scrutiny. The economics also matter, since running multiple commercial coding agents carries subscription and usage costs that scale with activity.

Still, the piece offers a concrete data point on where agent-assisted development is heading. It illustrates how a lone developer can plausibly expand their output by delegating implementation to AI while retaining the review authority that governs what actually ships. For engineers weighing similar setups, the described combination of Cloudflare Workers, a database layer, and two cooperating coding agents provides a reference architecture worth examining, alongside the honest reminder that human review remains the linchpin.