Claude Code 2.1.278, shipped September 19, added support for AGENTS.md -- an open, tool-agnostic convention for repo-level agent instructions that several coding agents (Codex, Cursor, Copilot, and now Claude Code) can read. The precedence rule is sensible and worth knowing: CLAUDE.md still takes priority when it exists, and Claude Code still looks for CLAUDE.md first -- AGENTS.md is what it falls back to, not a replacement for teams that already have a CLAUDE.md they like.
The practical unlock is for teams or open-source projects that support contributors using different agents: previously, standardizing repo-level agent instructions meant either picking one tool and writing for it, or maintaining near-duplicate instruction files per tool (a CLAUDE.md, a Codex config, a Cursor rules file) that inevitably drift out of sync with each other. A single AGENTS.md that multiple tools read directly closes that gap -- one file, one source of truth, whichever agent a given contributor happens to be running.
This lands in the same week as two other Claude Code changes worth noting together: version 2.1.270 (September 12) fixed a genuinely annoying bug where git commands would unexpectedly re-request permission mid-session on long-running work, and the agent view dashboard, custom themes, and effort-level controls have all shipped recently too. None of these are headline capability jumps -- they're the accumulation of exactly the "scaffolding, not just the model" pattern this digest keeps tracking, this time on the harness's own day-to-day usability rather than a new agent capability.
Competitively, Codex (0.155.1, September 19) and Claude Code are close enough on Terminal-Bench 4.0 this month (GPT-6 Astra at 58.2% vs. Claude Fable 5.1 at 57.9%) that the differentiator for most teams is genuinely shifting toward exactly this kind of workflow/interop polish -- AGENTS.md support, permission-prompt annoyances, dashboard usability -- rather than raw benchmark score, which is close enough between the top two that it's not the deciding factor anymore.
If you maintain agent instructions for a repo multiple contributors use with different coding agents, AGENTS.md is now worth adopting as the single shared file instead of maintaining per-tool duplicates -- and with the top two coding agents this close on raw benchmark score, harness-level polish (like this) is becoming the more relevant differentiator to evaluate.