Claude Code 2.1.277 reads AGENTS.md, so one file feeds every agent
Anthropic converges on the cross-agent config standard; the release also retires the TaskOutput tool, strips hidden Unicode from prompts, and speeds startup.
Copy markdown
One config file, every agent
Claude Code 2.1.277 now reads AGENTS.md as project instructions in any repo that has no CLAUDE.md, and /config lets you pick which file wins. Teams juggling separate instruction files for Claude Code, Codex, and Cursor can collapse them into the single AGENTS.md the ecosystem already shares. Bedrock, Vertex, and Foundry don't support it yet.
TaskOutput is gone — use Read
The release deprecates the TaskOutput tool along with the taskOutputMaxChars setting and the TASK_MAX_OUTPUT_LENGTH env var; subagent output now comes back through the ordinary Read tool, rendered as indented body text. Any script or plugin still calling TaskOutput will break on 2.1.277.
Invisible Unicode stripped before sending
2.1.277 strips invisible Unicode characters out of prompts before they reach the model — a direct block on the hidden-instruction prompt injections that smuggle commands through pasted text or fetched web pages. It matters the moment you pipe untrusted input into an agent.
Faster headless starts, fewer config crashes
SDK and headless sessions start faster, and a malformed ~/.claude.json (bad theme or claudeAiMcpEverConnected values) no longer crashes the CLI. Minor on paper, but exactly the failure mode that silently breaks CI and automation runs.