
What happened
The author tested Claude Code 2.1.170 and 2.1.278 in 12 tool-disabled runs on Windows; @AGENTS.md import returned both markers on both, but AGENTS.md alone returned NONE on 2.1.278.
Why it matters
The import path appears to be the one route that produced a marker across both versions, so the setup under test — not the feature itself — may be what decided the result.
What to watch
The lone-placement failure remains unexplained, since no internal load trace was captured and mode, tool settings and flag effects were not isolated. Watch the 2.1.277 native-support threshold.
WHO IT HITSDevelopers running Claude Code on their own machines and people who maintain shared instruction files in coding projects are the ones who would need to re-test their exact launch settings rather than assume version alone settles behavior.
Summaries like this, in your inbox every morning.
The test exists because the author wanted to separate the official specification from what actually happens on one machine. The current documentation places native support at 2.1.277 and later, with conditions tied to the presence of other instruction files, plus feature enablement and first-session constraints. Rather than assume an updated build behaves the same everywhere, he fixed the versions and compared unique marker strings in responses.
The 12 runs span six conditions each on 2.1.170 and 2.1.278, and every run logged an empty tool, MCP and skills set with no actual tool calls. That matters because it closes off the ordinary route where the model simply opens a file and later repeats what it read. Under those conditions the @AGENTS.md import produced both markers on both builds, while the explicit both-files mode produced a marker only on the new build.
The unresolved point is the lone-placement case on 2.1.278, which returned NONE even though the same build surfaced the AGENTS.md marker when both files were read. The author is careful not to call this a verdict on native support; he simply did not capture the internal load path, and interactive mode, normal tool-enabled runs, user settings and the effect of flags such as --disable-slash-commands were left untested. For teams relying on shared instruction files, the practical reading is that behavior may hinge on the exact launch configuration rather than the version number alone, and a small marker-string check before changing a live project looks like the safer path.
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 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.
ServiceNow launched Flow, a natural-language service desk that handles employee requests inside Slack and Micr…
DeepSeek and Huawei announced a partnership to develop semiconductor software, part of efforts to reduce China…

The AI Ataraxos beat Niemeijer, the most decorated Stratego player, with an 85 percent effective win rate over…

Google unveiled Gemini 4 Argon on September 30, saying it beats GPT-6 Astra and Claude Opus 5.5 on 13 of 19 be…

Yann LeCun told Fortune’s Emily Forlini he has zero concerns about rogue AI incidents, including OpenAI agents…

A Preply survey of over 5,000 professionals in nine countries found 92% of Gen Z respondents used AI for learn…
