KognitaKognita.

Blog

AGENTS.md Standardized the File. It Didn't Fix the Staleness.

10 min read

AGENTS.md is a genuine win. Instead of maintaining a CLAUDE.md, a .cursorrules, a .windsurfrules, and a copilot-instructions.md that all say the same thing, you maintain one file most tools agree to read. That solves a real, annoying problem. It also quietly reveals which problem the industry chose to solve — and which one it did not. Standardizing the filename does nothing about the contents going stale. We agreed on a format, not on a way to keep it true.

What it actually fixed

The fragmentation was real and the fix is worth having:

One file the tools agree on, instead of four copies
What AGENTS.md standardized

  Before:  CLAUDE.md      (Claude Code)
           .cursorrules   (Cursor)
           .windsurfrules (Windsurf)
           copilot-instructions.md (Copilot)
           -> same content, copied into N bespoke files

  After:   AGENTS.md       -> one file most tools agree to read
           -> real win: stop maintaining four copies of the same prose

  Standardized: the FILENAME and FORMAT. Not the freshness.

No more copy-pasting the same conventions into four bespoke files and watching them diverge. For teams using multiple AI tools, that alone justifies the move. But notice exactly what got standardized: the location and the format. The thing inside the file is unchanged — it is still hand-written prose describing a codebase.

What it left exactly as broken

Every structural weakness of CLAUDE.md survives the rename:

A shared static file is still a static file
What standardization left untouched

  A shared static file still:
    -> is hand-written prose about code that keeps moving
    -> has no owner, no update trigger, no expiry
    -> goes stale the moment the codebase changes
    -> "a stale entry is worse than a missing one — it
       resolves ambiguity in the wrong direction"

  One agreed-upon filename, four tools, same drift.

As one practitioner put it, "when a convention changes, the old directive actively harms the agent's output" — and "a stale entry is worse than a missing one, because it resolves ambiguity in the wrong direction." That is true whether the file is named CLAUDE.md or AGENTS.md. It is the same drift documented in why static context files can't keep up, now blessed with a standard.

Standardizing the file is not standardizing the truth

There is a meaningful difference between every tool reading the same file and every tool reading the same source of truth:

An agreed file vs. an agreed source
Standardizing format vs. standardizing source

  AGENTS.md:
    -> every tool reads the same FILE
    -> but the file is still a snapshot someone typed
    -> freshness is still a human chore

  Shared semantic index:
    -> every tool reads the same SOURCE OF TRUTH
    -> derived from code, re-derived on change
    -> freshness is the system's job, not yours

A standard file means consistency across tools at a single moment — and identical staleness across all of them a month later. If the AGENTS.md is wrong, it is now wrong in Claude Code, Cursor, and Copilot simultaneously. Standardization propagated the convenience and the drift equally. (A small but telling detail: Claude Code does not even read AGENTS.md directly — you import or symlink it — so the "one file" promise already has asterisks.)

The thing worth standardizing

If the goal is for every tool to share one consistent picture of the codebase, the right shared artifact is not a file someone types — it is an index derived from the code itself, queried through an open protocol like MCP. Then every tool reads from the same current representation, and "keeping it true" is a property of the system rather than a recurring chore replicated across however many tools adopted the standard. The format question was always downstream of the freshness question.

Where Kognita fits

Kognita is the shared source of truth that a shared file only imitates: a semantic index derived from your repositories, re-indexed as they change, exposed to any MCP-capable tool. Whatever editor or agent your team uses, it grounds in the same current representation of the code — not a snapshot someone last edited before the refactor. Keep AGENTS.md for stable cross-tool conventions; let an index carry the codebase facts that a standardized file standardizes the staleness of.

Final take

AGENTS.md is a good answer to "why am I maintaining four copies of the same instructions." It is not an answer to "why do my instructions go stale," because that was never a format problem. The industry standardized the file and left the freshness exactly where it was.

One agreed-upon filename does not make the contents true. Standardize the source the tools read from — a current index — not just the file they read it in.