Claude Code adds AGENTS.md support and drops auto mode classifier fees
Claude Code now looks for AGENTS.md for project instructions and stops charging for auto mode classifier overhead.
Tools · Source: Hacker News
What happened
Anthropic just shipped Claude Code versions 2.1.277 and 2.1.278. The biggest change is how the CLI handles project instructions. If a repository lacks a CLAUDE.md file, Claude Code now automatically reads AGENTS.md instead. You can configure this under project instructions in the settings. This feature is live for direct users but not yet available on Bedrock, Vertex, or Foundry.
The update also overhauls auto mode billing. Claude API, Enterprise, and cloud provider users now default to a server-side classifier. This change eliminates the billing overhead previously charged for classifier routing. Users get a new status row showing if the session runs on the server. You can opt out using a specific environment variable if needed.
Anthropic patched massive edge cases across the board. Subagent results now reach the main agent indented and explicitly marked as subagent output. The deprecated TaskOutput tool is completely gone. Claude now uses the standard Read tool for background task outputs. The VS Code extension got heavy updates too. Builders get a new session cost tracker, background shell management, and one-click copy buttons for responses. Gateway users get new egress boundary headers. Every outbound request can now hand a proxy the hostname instead of resolving it locally. Anthropic also fixed a critical prompt cache miss that happened when resumed sessions printed output from a SessionStart hook.
Key facts
- 2.1.278 — The latest Claude Code version released
- CLAUDE_CODE_AUTO_MODE_SERVER=0 — Variable to opt out of server-side classifier
- AGENTS.md — New default file for project instructions
- CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY=1 — New header for Claude apps gateways
Why it matters
Standardizing on AGENTS.md signals a major shift. We are moving away from generic AI coding assistants toward specialized, project-aware agents. You no longer have to cram agent instructions into a generic text file meant for human developers. You can isolate machine rules completely. This creates a cleaner repository structure for teams building complex software.
Dropping the classifier overhead cost is a quiet but massive financial win. When you run agents on auto mode across thousands of tasks, routing fees destroy your margins. Defaulting to the server-side classifier makes autonomous coding radically cheaper at scale. It removes the penalty for letting the model decide how to route tasks.
The security and context isolation changes matter just as much. Subagent outputs are now strictly formatted and indented. This prevents text in a subagent result from passing as the main session instructions. It is a direct mitigation against prompt injection and context confusion. Furthermore, fixing the prompt cache misses on resumed sessions means you stop paying for redundant token processing when you restart a complex workflow. You get faster resumes and lower bills.
For builders
Migrate instructions to AGENTS.md
Move your repository agent rules from CLAUDE.md to AGENTS.md. This keeps human and machine documentation separate. Bedrock, Vertex, and Foundry users must wait for support.
Check your auto mode billing
You no longer pay for classifier overhead by default. If you need the old behavior on gateways, opt out using the CLAUDE_CODE_AUTO_MODE_SERVER environment variable.
Update your task output tools
The TaskOutput tool is officially dead and removed. If your custom scripts relied on it, they will fail. Switch to using the standard Read tool for background task outputs.
Audit your bash sandbox rules
Anthropic fixed a bug where a single matched glob exempted an entire compound Bash command from the sandbox. Every part of a compound command must now match your exclusion rules.
My take
Anthropic is quietly building the operating system for agentic coding. Moving to AGENTS.md proves they know developers want multi-agent orchestration, not just a smart autocomplete. Dropping the routing tax and fixing prompt cache misses just accelerates the inevitable adoption of fully autonomous dev teams.
Original reporting: Hacker News. This is my rewrite and opinion.