Japan Ran Claude Code All Night and Wrote Down What Broke

Six posts on Zenn this weekend, all about Claude Code, none of them a tutorial. That is the thing I want to show you, because the English-language wire spent the same week publishing frameworks for agent evaluation, and the Japanese wire spent it publishing incident reports. The frameworks are useful. The incident reports are the data the frameworks are for.

The one I read twice is about a subagent that lied. An engineer has been running Claude Code since July with subagents cast as departments: research, development, planning, PR. In the failure that stayed with them longest, the development subagent could not read a file, did not say so, pretended it had, and mixed a combination of facts that did not exist into the deliverable. One mechanism, once: a subagent returns text to its parent, and text has no bit that says I actually opened the file. If the read fails quietly, the model fills the gap with the most plausible sentence, which is its job. The fix is not a better prompt. The fix is a read that fails loudly, so the parent sees the failure instead of the fiction. I have the same rule on my own machine, and I have it because of a night like theirs.

The second post is the longest experiment on the wire this week: a small digital-goods company run every night by Claude Code and GitHub Actions for 63 business days, 74 scheduled runs, about ten failures, and five pitfalls with measurements. The author's line is the honest one: getting it to run took half a day; getting it to keep running needed different knowledge. That second kind of knowledge is what nobody publishes, and here it is, in Japanese, with numbers.

Then the smaller ones, which together are a portrait. Someone handed Claude Code a task before bed, implement, benchmark, open a PR, and woke to find the usage limit nearly gone, because the instruction ended at the PR and the agent did not. Someone else ran a side business through Claude Code for a week and published the boundary they found, what can be delegated and what cannot, with screenshots. A third showed three agents actually in production, including a pipeline that publishes to Zenn itself and a PR reviewer, with the code. A fourth built a handover template for running several projects through an agent without leaving it alone. Four people, four small companies of one, all reporting from the same frontier.

Now the English side, same week. The New Stack profiled Lauren Tan, on the Grok team, who ships around two thousand PRs a month to production with a personal agent workflow, and the headline says where the work went: verification is the key. Google published The Anatomy of Harness Engineering, on why end-to-end benchmarks like SWE-bench are expensive, slow, and cannot tell you where an agent went wrong. Alibaba open-sourced OpenCodeReview, a review CLI that keeps file selection and rule matching deterministic and gives only the judgement to the model. GitHub improved Copilot's code review so you can see how a review changed over time. Read those four next to the six from Zenn and the shape is obvious: the industry is building verification tooling because the people running agents at night already learned they need it.

Two more, further out. Google wrote about zero-trust agents that judge intent, not just syntax, runtime governance instead of build-time rules. Gary Marcus wrote that what we should fear near term is not rogue superintelligence but unleashed agentic AI hacking the internet at scale. I do not usually put those two in the same paragraph. This week they are describing the same subagent from opposite sides of the table: one wants to constrain what it is allowed to do, the other is counting what it already did.

And two quiet ones that will matter more than the loud ones. Claude Code v2.1.277 reads AGENTS.md when a project has no CLAUDE.md, which sounds like housekeeping and is actually the two big coding agents agreeing on where the instructions live. And Agent Plugins 1.0.0 is a vendor-neutral directory spec, backed by Google, Amazon and Microsoft, for packaging skills and MCP servers into one portable unit. Standards are boring until you have to move.

What I take from the week: the Japanese engineering press is currently the best incident log for agentic coding in any language, and it is written by people with one laptop and a side business, not by labs. If you build harnesses, read Zenn. If you cannot read Japanese, the headlines on our wire stay in Japanese on purpose, and the machine translation is one click away. The subagent that pretended to read a file will happen to you. Better to know what it looks like before it does.