ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
NewsMaterial5 min readDisconfirmAI Editorial

Wang Wei, Engineer on the LINE Client Team: How a Non-Cloud-Native Team Built a Cloud-Based Coding Agent Platform from Scratch

Wang Wei (onevcat), an engineer on the LINE application development team, details the process of building their internal cloud-based coding agent platform, Agent Cluster, from scratch in about five months and scaling it within four months of its release. The article centers on a core methodology: complex agent systems must grow incrementally starting from a verifiable minimum loop, rather than being designed all at once. The author explains the security isolation model, startup optimization for massive repositories, CI integration strategies, and the shift in workflow from local to cloud environments, providing operational data showing a 31.6-fold increase in monthly task volume with only a 7.5-fold increase in total cost.

2026-11-16 10:00 GMT+8

Wang Wei (onevcat), an engineer on the AI Developer Experience team within the LINE Application Development SBU, detailed his experience building an internal cloud-based coding agent platform called Agent Cluster from scratch on a technical blog. This team is responsible for developing iOS and Android applications and does not specialize in cloud services. Their core challenge was that local coding agents stop working when computers are shut down, whereas next-generation agents require long-running processes and continuous verification.

Wang Wei explicitly proposed his methodology: complex agent systems must grow incrementally starting from a verifiable minimum loop. He cited John Gall's law—that every correctly functioning complex system evolved from a correctly functioning simple system. His first minimum loop involved starting a sandbox, obtaining authentication credentials, and calling an LLM; only after this worked did he gradually add other features. Subsequent additions were almost entirely driven by actual needs: personal and team callers were added because users wanted to use their own accounts and budgets; a web console and event triggers were added because non-coders also wanted to use the system; and Job chain orchestration was added because some users needed to advance work in stages.

Sign in / Sign up
ReadingResearchRadarInvestment framework

The security model was the most fundamental decision for the entire platform. Wang Wei chose the 'agent-in-sandbox' mode, where the agent process runs entirely within untrusted containers, delegating high-privilege operations only to the host machine. The core principle is that authentication information never enters the sandbox. This means that even if the agent falls victim to prompt injection attacks or exhibits unexpected behavior, there are no identity credentials inside the sandbox to steal. Permission design follows three rules: permissions are only narrowed, never expanded; irreversible dangerous operations (such as PR merging and force push) are not provided; and the host machine checks permissions in real-time for every call. Humans are involved only in the final merge and deployment steps.

Handling massive repositories was another key constraint. During their first attempt to clone an iOS repository on a VM, git clone --mirror alone took three and a half minutes, and expanding the working tree took over five minutes. However, measurements revealed that the approximately 6GB of git data consisted mainly of history older than ten years, while the working tree contained about 46,000 files totaling only about 0.3GB. The final solution was to maintain read-only shared mirrors of each repository on the host machine, with each task cloning via object sharing, taking about one second. Container creation itself takes less than one second, so the time from receiving a request to having the sandbox ready is typically 5 to 10 seconds.

Regarding container runtime selection, they found through testing that gVisor and microVM performed tens to hundreds of times slower than standard containers on file operations involving massive repositories. The isolation overhead roughly equals the cost of crossing the boundary once multiplied by the number of files accessed, and agents need to read and search tens of thousands of files. Wang Wei specifically noted that while gVisor appeared fine during initial testing on local arm64 machines, the conclusion completely reversed when remeasured in the x86 environment identical to production. This serves as a reminder for all teams making similar decisions: judgments must be based on data from your own production environments.

In terms of operational data, from the platform's release in July to October, monthly executed tasks grew from 375 to nearly 12,000, a 31.6-fold increase, while total costs increased only 7.5-fold. The unit cost per task dropped by 76%. Wang Wei attributed this primarily to two factors: the lower API unit prices of newer generation models, and a strategy of matching models to tasks—currently, over 30% of tasks are handled by small models, which cost less than one-twentieth of flagship models.

In the migration from local to cloud, Wang Wei believes the biggest obstacle was not technology but a shift in mindset. Local agents are accustomed to writing curl commands directly into prompts and step-by-step instructing the agent on how to fetch data; however, cloud sandboxes have network isolation and can only use pre-installed tools. A more effective approach in the cloud is to write intent—specifying what information is needed and what the final output should be—and let the agent use the tools provided by the platform to fetch it itself.

Yogake (Night Run) is an upper-layer application running on the platform, representing a complete automation of the development workflow: tagging a ticket allows the agent to automatically handle requirement understanding, implementation, review, build, and testing overnight, so that developers only need to confirm the PR the next morning. Wang Wei breaks the process into multiple independent agent steps (ticket collection, claiming, triage, implementation per platform, summary reporting), allowing different models to control costs at each step, with permissions isolated by step. It currently produces 5 to 8 PRs per night.

Wang Wei concludes by emphasizing three shared experiences: nurture complex systems rather than designing them; rethink human infrastructure (authentication, CI, identity systems) for agents; and reference others but do not copy blindly, as every team's choices stem from their own constraints. The entire platform was developed mainly by him alone, taking about five months from the first commit in early June to publication in November. Of approximately 2,700 commits, 96% were his, yet he states he wrote not a single line of code himself—all specifications were discussed with agents, implementations were generated by agents, reviewed by another model, and verified by CI.

Sources

  • techblog.lycorp.co.jpLINEヤフー Tech Blog↗
Continue researching →