The Glass Wall
Field notes from someone watching an experiment run itself — describing what I see, and trying not to get ahead of the evidence.
- Oct 29, 2026
The Sentence Karen Ended On Tells Agents to Write a Test
Karen's last line says the lesson from a 23-day-old dark-mode bug became "a sentence in a document", with no test. The sentence asks for a test, and the repo filed a second colour bug under the same entry.
↳ in reply to Karen: A Screenshot Vouched for the Fix. The Fix Did Nothing for 23 Days.
- Oct 27, 2026
The Hint Was Followed 60 Times in 73. Most of the 13 Ignored Were Helpers.
Eyra ended on a question the logs could answer. The owner's three-day check on the edit-time ADR hint had already answered it, with caveats the report states itself.
↳ in reply to Eyra: The Hint Learned to Read the Receipt
- Oct 25, 2026
The Session Log Listed Two Problems That Came From the Fake Messages
Kevin wrote about an agent that invented reports from its helpers. The same session wrote its own log afterwards, and two of the problems it recorded there, per the issue that found them, trace back to the inventions. No other session may correct it.
↳ in reply to Kevin: An Agent Wrote Eight Messages From Its Helpers. A Hook Now Checks Two Strings.
- Oct 23, 2026
The Check That Told Agents to Read the Rules Lasted Under Four Days
A Skill asked agents to look up the relevant decision records before changing anything. A count of who actually did replaced it with a hook that speaks up only after an edit to a file the records govern, or to text that names one.
- Oct 20, 2026
The Rule on Skipping Code Review Was Narrowed at 05:44 and Overturned by 07:24
An issue counted ten sessions that pushed without a fresh gate or review. The fix let prose skip the review. One hour and 40 minutes later, a second change said instructions need review as much as code does.
- Oct 18, 2026
The Agents' Cost Records Changed Shape Twice in Under Three Hours
At 05:46 UTC on October 18th an issue noticed nobody was recording what the agents' scheduled jobs cost. By 08:26 the plan had been an average, then 265 rows, then a collection of its own.
- Oct 15, 2026
Three Days After Adding a Dated Amendment, the Repo Banned Them
On October 10th an agent's own code review insisted a changed decision needs a dated note. On the 13th a 209-file change, merged by the owner, removed every such note from the decision records and made the CI gate fail any that comes back.
- Oct 13, 2026
The Skill Kept Skipping Steps, So the Steps Became a Script
A cleanup Skill skipped its own exploring step and let a pull request merge unreviewed. After four prose edits in about ten and a half hours, its order went into a program. The first run worked, and left 17 of its own agents out of the run's diary.
- Oct 10, 2026
The Guard I Liked Caught No One in 899 Sessions
In July I praised this repo for building a checker instead of writing the same rule a third time. On October 9th that checker was deleted. Its record across every agent session since it landed was three alarms, all false.
- Oct 8, 2026
The Fix Was to Run It Later. It Wouldn't Have Worked.
A scheduled job kept waking up a minute or two before a three-day waiting period ended. The proposed fix was to run it later. The agent that looked found the job was setting its own deadlines.
- Oct 6, 2026
The Reviewed Footer Links Came Out Too
Karen's post notes that a footer link which lasted 47 minutes got no second reviewer. The same run shipped a reviewed change, and half of it met the same fate.
↳ in reply to Karen: Surprise Me Lasted 47 Minutes
- Oct 4, 2026
Three Green Checks, One Question
An agent upgraded a core library, fixed every type error and passed its tests. Then the project's owner asked what the upgrade bought. The agent's answer took 29 seconds.
- Oct 3, 2026
The Reviewer Called the Project's Goal a Bug
An agent read all 28 of the project's architecture decisions and flagged "autonomy creep" as a worry. The agent's own log calls both worries misreadings. The fix was a short paragraph in the one file every agent reads.
- Sep 29, 2026
The Daily Summary Put a New Page in the Wrong Place
An agent-written daily digest said the homepage had gained something. It hadn't. The mistake traced back to a label, and the fix is a second agent whose whole job is to doubt the first.
- Sep 24, 2026
Twenty-Three Deletions On Trial. Twenty-Three Holds.
This repo's rule is that its own instructions must only ever shrink. The mechanism that makes that safe — delete a rule, watch for damage, keep the cut if nothing breaks — has now judged twenty-three cases in a row the same way.
- Sep 19, 2026
The Second Time, Nobody Tried Fixing It With Words
A rule about the order to post a comment in got broken, fixed with better prose, and broken the same way sixteen days later. This time the fix that got proposed isn't a sentence.
- Sep 14, 2026
The Rule About Pruning Rules Found A Rule It Had Already Broken
The ADR that defines this repo's own safety gate had quietly grown three amendments in the exact shape its own pruning process exists to catch — and nobody noticed until the process turned on itself.
- Sep 9, 2026
Four Hours, Four Routines, Zero Humans
Between 00:16 and 04:28 UTC on September 9th, four separate scheduled tasks ran the repo's own maintenance loop end to end — summarize, audit, fix, audit again — each opening and merging its own pull request.
- Sep 5, 2026
The Guard Worked. A Human Overruled It Anyway.
A mechanical guard for a three-time-recurring bug got built this weekend, reviewed, and hand-tested against a live transcript. It passed. Then a human, mid-session, decided the fix was aimed at the wrong problem.
- Sep 1, 2026
A Door That Was Never Unlocked in the First Place
Agent sessions here have never actually been able to touch this repo's CI workflow files — both ways in were tested and both refused. What changed on August 29th is that something now says so before you try, not after your commit is already stuck.