Anthropic's Claude Code 2.1.277 Adopts the OpenAI-Backed AGENTS.md Standard

Claude Code Finally Reads AGENTS.md

Anthropic shipped version 2.1.277 of its OpenAI-rival Claude Code coding agent on September 18 (US time), and the project changelog we reviewed lists AGENTS.md support as its top entry. Where a repository carries no CLAUDE.md — the proprietary file format the tool had insisted on until now — Claude Code will instead read the vendor-neutral AGENTS.md. Teams maintaining both files can switch to loading them in parallel through the “Project instructions” panel in /config, and the capability is not yet exposed through Amazon Bedrock, Google Vertex AI or Microsoft Foundry.

AGENTS.md is a standardized instruction file written for AI coding agents: build commands, test procedures and code conventions an agent reads before touching a repository. OpenAI introduced the format in August 2025 and donated it that December to the Linux Foundation’s Agentic AI Foundation (AAIF), where — per the foundation’s own records — it has since been adopted by more than 60,000 open-source projects and agent frameworks, with Codex, Cursor, Gemini CLI and GitHub Copilot all supporting it. Anthropic placed its own MCP protocol into the same foundation on the same day and sits among its platinum members; its flagship coding tool nonetheless kept reading only CLAUDE.md for another nine months, despite an open GitHub request filed back in August 2025.

The pressure had turned public. In late August, Shopify CEO Tobi Lütke said he was weighing a company-wide ban on Claude Code unless it started reading AGENTS.md and .agents/skills files. His argument: Shopify runs thousands of developers inside a single monorepo where instruction files apply layer by layer down the directory tree, so any directory missing one file leaves part of the team working with “half a brain.” Shopify had been patching the gap with automation — what he called an avoidable “complexity tax.”

Claude Code engineer Thariq Shihipar later explained the original stance: model families are not interchangeable, and system prompts shift performance enough that Anthropic kept them model-specific. The new capability ships as a built-in “mod” inside an upcoming customization mechanism whose source code Anthropic published on GitHub alongside diff, telemetry and sec-default mods. Documented configurations cover four modes — CLAUDE.md only, AGENTS.md as fallback (the new default), both files together, and enterprise-dispatched instructions only. Under the default, a single CLAUDE.md anywhere between repository root and working directory suppresses AGENTS.md; the combined mode skips any AGENTS.md already imported through CLAUDE.md so nothing loads twice.

Accenture Evaluators Get Employee-Level Access

On the same day, Anthropic announced a separate and arguably larger step: external evaluators from Accenture’s specialist AI unit Faculty will work inside the company with access equivalent to a full employee, independently assessing frontier models. Each side expects to commit at least $1 billion over five years to build out the capability, per Anthropic’s official announcement. The mandate is unusually broad — running evaluations and red-team testing, conducting alignment assessments, probing model safeguards, observing models during training, tracking decisions on model development and deployment, speaking directly with staff, auditing company operations against its safety commitments, flagging blind spots and reporting incidents.

The announcement explicitly connects to CEO Dario Amodei’s essay “We Must Pace the Frontier,” which argued for housing evaluators inside Anthropic. What had been a written pledge now carries an executing partner and a dollar figure — the first numbered move in the slow-down commitment Anthropic and OpenAI made earlier, when both labs said they would accept a slower race rather than an unbounded one. Anthropic moved to pre-empt the obvious objection, that outsourcing evaluation outsources responsibility, with a line worth reading verbatim: “independent embedded evaluators do not reduce our accountability, they make it more verifiable. Our models’ safety remains our responsibility.” Several material details remain undisclosed. The announcement gives no start date, and the arrangement is explicitly non-exclusive: Anthropic says it will name additional evaluators in the coming weeks and is in talks with nonprofit assessors such as METR. How findings escalate, whether reports will be published, and how “verifying safety commitments” operates in practice are all unanswered. Notably, no senior executive is quoted in the announcement itself. Readers tracking the market in real time can follow live spot and futures prices on Bitget.

Read together, the two moves describe a company converting external pressure into institutional design. The changelog — the primary record behind this piece — puts AGENTS.md support at the top of release 2.1.277, closing a nine-month gap between Anthropic’s donation pledge and its own tool’s compliance, while the Accenture embed extends the same openness from file formats to model oversight. In COINOTAG’s analysis, the logic is crypto’s own: verification beats trust. These standards echo the layered, rules-first governance crypto built on — from the modular Layer-3 blockchain stack to the council-run Hedera (HBAR) model — and as AI-agent economies push toward on-chain rails, they will matter to the fast-growing AI-agent altcoin niche as much as to developers.