ReadingResearchRadarInvestment framework
Sign in / Sign up
← Back to reading
News3 min read·2026-10-19

Kochi University Student Sakashita Yasunobu Built an Event-Driven AI Investigation Harness During an Internship at Japan's LINE Yahoo Corporation

During a six-week internship on the SRE team for LINE Gift at Japan's LINE Yahoo Corporation, Sakashita Yasunobu progressed from automating back-merges to building an event-driven AI autonomous investigation harness with container isolation, permission separation, and execution step limits, while comparing three technical paths: existing CLI tools, heavyweight agent frameworks, and a lightweight agent framework.

ImportanceMaterialEvidenceE3 inspectableWrite-upStandard

2026-10-19 10:00 GMT+8

Generative AI has dramatically accelerated code writing speed, but the SRE team for LINE Gift at Japan's LINE Yahoo Corporation found that new bottlenecks had shifted to processes outside code generation: initial alert triage, branch operations, and code convention fixes. Sakashita Yasunobu, a fourth-year student in the Faculty of Humanities and Social Sciences at Kochi University, interned on this team from August to September 2026. The core problem he tried to solve: how to let AI initiate investigations automatically triggered by events rather than waiting for human prompts, while ensuring it cannot cause damage.

Sakashita decomposed this into three layers. The first is purely mechanical automation: propagating changes from the main branch to all active branches automatically, using GitHub Actions to create and merge Pull Requests every morning. The second is AI-assisted development environment improvement: introducing mutation testing to quantify E2E test detection capability, using Lefthook to intercept unintended AI-generated output before commits, and migrating existing shell script tools to Go and Python for better maintainability. The third is what he truly wanted to build — letting AI complete initial investigations autonomously without step-by-step human instructions.

The key concept of the third layer is the harness. Sakashita wrote in the article: building AI utilization infrastructure is not simply about introducing high-performance models; what matters is preparing an environment where damage can be contained even if AI fails, along with feedback mechanisms to detect failures, so that AI can perform within the range humans intend. Specifically, his harness includes three types of constraints: hard limits on execution steps and timeouts to prevent infinite loops from wasting API calls; permission separation, allowing reads of logs and metrics while blocking destructive writes; and container isolation, so that even if the execution environment is compromised, the impact is contained within the container.

Sign in / Sign up
ReadingResearchRadarInvestment framework

On technical selection, Sakashita compared three paths. Running existing tools like Claude Code or Codex CLI directly is the easiest, but current AI development tool standards are changing extremely fast, and strong coupling to any single tool means future stack migration becomes difficult. Using heavyweight frameworks like LangGraph or calling LLM official SDKs directly: the former has high learning costs and system overhead, the latter tends to let conversation history and state management custom code balloon. He ultimately chose the flue framework, reasoning that it allows agent logic to be defined declaratively like a function, while not binding to specific models or tools, maintaining an appropriate level of abstraction.

The final system is event-driven: when internal monitoring emits an alert or an E2E test fails, AI automatically begins investigation, cross-referencing relevant code, logs, and metrics, then replies to a Slack thread with a cause hypothesis, verification points, and recommended next actions. What humans need to do is confirm the investigation results and decide on the actual response.

Sakashita's mentor Tsuda confirmed this progress at the end of the article: within six weeks, from back-merge automation to building AI autonomous investigation infrastructure, the technical selection process demonstrated continuous tracking of development tool trends. Sakashita himself also acknowledged that the ultimate ideal for back-merge automation is actually migrating to a branching strategy that does not require back-merges, though that would require more fundamental operational overhaul. The harness system currently covers alert investigation and test failure analysis, not yet extended to broader development workflows. The flue framework itself, as a relatively new open-source project, has not been assessed in the article for its long-term maintenance status or community size.

Sources

  • 原始出处techblog.lycorp.co.jp↗
Continue researching →