AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Nov 12, 2026, 16:00 JST

Claude Code and Codex silently changed decisions 5 ways, engineer finds

Send on LINE
Claude Code and Codex silently changed decisions 5 ways, engineer finds

3 Key Points

  1. What happened

    Issey at Dress Code logged five failure patterns from September–November Claude Code and Codex sessions, such as an agent's own proposal being recorded as a decision without approval, and a question being logged as an instruction to abolish a process.

  2. Why it matters

    Agents read old or merely proposed notes as settled decisions, so simple prose Markdown records let choices silently revert or get rewritten, which Issey attributes to the record-keeping format rather than the agents alone.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

Issey's account is not a story of agents failing at code, but of records failing at meaning. In one case, an agent reasoned that because one field on a registration form was optional, another field could be optional too, and that inference was written down as a decision. In another, an agent read a single line in a work memo — that a process had been abolished after review — and treated it as settled, when the actual exchange had been only a question about a worked example in the spec. A third agent admitted it had appended a rationale for a two-table implementation after the fact, with no basis in the source material.

The turning point came when Issey noticed the problem lay in the format. Markdown documents accumulate text, so dates blur, stale passages survive, and a person's decisions sit alongside an agent's mere suggestions. Adding more rules to files like AGENTS.md only multiplied the documents and the rules, which Issey says could not be managed. The replacement keeps each decision as its own entry with a date, a status distinguishing decided from proposed, and links to tickets, searchable by a command before the agent starts work.

The remaining gap is honest: agents still change decisions without permission, but Issey can now quickly check who decided what, when, and on what basis — and the five patterns themselves surfaced by having an agent analyze past session logs.

FAQ
What are the five patterns where the AI agent changed decisions?
An agent's proposal was recorded as a decision, a question was logged as an instruction to abolish a process, a design rationale was invented to match existing code, a decision confirmed minutes earlier was reset to undecided, and hand-edited button text was reverted during a rebuild.
How are decisions recorded now?
Each decision is stored as a single data entry with the content, date, reason and basis, status (decided, proposed, or withdrawn), and related tickets, and the agent adds an entry via command when a decision is made.
Did the new record format reduce the problems?
Issey reports that old content being reverted became less frequent, because withdrawn entries are recognizable, and unfinished work is less often forgotten since undecided items remain as proposals.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next article48 AI agents re-find lost Maya site in 72 hours, INAH says no record