Anthropic Adds AGENTS.md Support and Parallel Projects to Claude Code

Anthropic has introduced support for OpenAI's AGENTS.md specification in Claude Code, starting with version 2.1.277. [2][3] Claude Code engineer Thariq Shihipar announced the update in a social media post. [2] The tool will now automatically scan for and load an AGENTS.md file if a CLAUDE.md file is absent in a given directory. [3][7] Users can toggle this behavior using the /config command. [2][3]

Previously, developers using Claude Code alongside non-Anthropic tools had to maintain separate CLAUDE.md and AGENTS.md files. [2][3] These Markdown documents are read by coding agents with every request to establish context, define coding conventions, and outline preferred tool usage. [2][3] To keep the two instruction formats synchronized, some engineers relied on workarounds like creating symlinks. [2][3] OpenAI contributed the AGENTS.md format to the Agentic AI Foundation under the Linux Foundation last year, and over 60,000 open-source projects had implemented it by December 2025. [2][3] Following the release, OpenAI's head of core product Thibault Sottiaux posted, "Yay! This is the way. Come to the light." [2][3]

The requirement to maintain separate instruction files created workflow conflicts that were explicitly highlighted by Shopify CEO Tobi Lütke. [4][7] Failing to synchronize these rules posed a risk of conflicting AI instructions across complex enterprise codebases. [4][7] By addressing these concerns, Anthropic secured continued usage among major enterprise accounts that were previously at risk of abandoning the tool. [4] Anthropic implemented this capability by making AGENTS.md a built-in "mod" and publicly releasing the corresponding source code. [8][9] Thariq Shihipar noted that developers will eventually be able to build custom versions around their project instructions using this mod system. [8]

However, the AGENTS.md integration is not yet available on major cloud platforms like Google Vertex and AWS Bedrock, delaying its use in some enterprise infrastructure. [4] Additionally, the AGENTS.md file is not read during the very first session after introduction, only becoming effective from the second session onward. [10] If a file named CLAUDE.local.md is present, the AGENTS.md file will not be loaded, requiring users to modify their settings.json if they want both formats read. [10]

Alongside the configuration changes, Anthropic has overhauled the Projects feature in Claude Code to allow Pro and Max subscribers to run multiple parallel sessions on a single codebase. [5] Each task operates as its own cloud session with a dedicated repository copy and git branch. [5] A coordinator thread directs the parallel workers, and when threads touch the same code, merge conflicts appear similar to standard pull requests. [5] Claude Code now provides five parallel execution modes, including Dynamic Workflows for orchestrating subagents with scripts and Agent Teams for coordinating instances that share information. [5]

Anthropic is framing this redesigned Projects tool as a coding "chief of staff" that tracks how one assignment affects another to prevent changes from quietly breaking related tasks. [6] The company warned that because Projects run multiple full Claude Code sessions simultaneously, users can reach their usage limits faster. [5] In a 2026 Sonar survey of over 1,100 developers, 96 percent reported not fully trusting AI-generated code, with 38 percent noting that reviewing AI code took longer than reviewing a colleague's work. [6] This parallel execution feature is currently in beta, and users without existing Projects can join a waitlist, though Team and Enterprise plans remain excluded. [5]

Anthropic's decision to embrace an OpenAI-originated standard signals a shift away from specification separatism in the competitive AI coding market. By supporting AGENTS.md, Claude Code simplifies the workflow for teams concurrently using rival products like OpenAI Codex, GitHub Copilot, and Cursor. The parallel execution capabilities in Projects also arrive as developers rely on tools to coordinate larger workflows, competing with platforms like Cognition's Devin and Amazon's Kiro. Furthermore, the introduction of the "mods" framework suggests Claude Code is moving toward a customizable architecture, contrasting with the "everything is a plugin" approach of the recently open-sourced DeepSeek Harness. What the sources don't address: when Anthropic expects to bring AGENTS.md compatibility to enterprise cloud platforms like AWS Bedrock and Google Vertex.