CONFIRMED: Claude Code 2.1.277 expands project instructions and hardens agents

Anthropic released Claude Code 2.1.277 on September 18, Brasília time, in an update that looks like maintenance but touches the foundations of daily operation. The version reads AGENTS.md when a project has no CLAUDE.md, adds controls for gateways behind a proxy, and fixes failures that could stall headless sessions, plugins, resumes, and agent workflows. There is no new model here. The value is a less fragile environment when the tool becomes infrastructure.

A more open instruction convention

AGENTS.md reduces dependence on a convention exclusive to Claude Code. In projects without CLAUDE.md, the file becomes an instruction source, with the setting available under Project instructions in /config. The limitation matters: the release says this behavior is not yet available on Bedrock, Vertex, or Foundry. For a small team, this makes it easier to share review rules, coding standards, and tool boundaries across agents, but every execution environment still needs its own test.

The gateway becomes an explicit boundary

Two features are especially useful for operations that centralize traffic. CLAUDE_GATEWAY_PROXY_IS_EGRESS_BOUNDARY tells the gateway that a forwarding proxy is the egress boundary, so each request gives the proxy the hostname instead of resolving it locally. The release also adds an optional static headers map to gateway upstreams. That helps integrate authentication, routing, and observability without placing provider specific logic inside every project.

Fixes that matter in production

The long list of fixes matters more than any flashy item. claude -p sessions and Agent SDK sessions that could end without a result after an internal error now report the failure and exit with code 1. The release also fixes resumes involving empty text blocks, unexpected logouts when an older build runs on the same machine, crashes caused by malformed configuration, and Grep or Glob searches that hid process, memory, or file handle exhaustion.

Security and compatibility are part of the release

Prompts containing invisible Unicode characters are cleaned and shown for review before sending. Subagent results now arrive clearly marked as subagent output, reducing the chance that external text will be mistaken for an instruction from the main session. TaskOutput was removed, and workflows are expected to read the output file with Read. That is a compatibility change worth checking in custom automation.

What a small business should do

A small or midsize company should update in staging if it uses the Agent SDK, headless execution, gateways, plugins, or subagents. The minimum test should open a project with AGENTS.md, run a noninteractive task, resume a session, and validate the proxy path. Teams should also search scripts for TaskOutput references. Simple environments without these features do not gain a product changing capability from this release.

Reliability over hype

My reading is positive, without marketing fireworks. This is a reliability release: it reduces silent failures, clarifies the boundary between agents and infrastructure, and prepares Claude Code for more distributed operations. The counterpoint is the breadth of the patch. Automatic updates without staging remain a bad idea. The npm registry confirms package 2.1.277 was published in the same cycle, while the official changelog supports the CONFIRMED classification.

Sources

Anthropic official GitHub release: https://github.com/anthropics/claude-code/releases/tag/v2.1.277 | Official structured release record: https://api.github.com/repos/anthropics/claude-code/releases/tags/v2.1.277 | Independent npm package registry record: https://registry.npmjs.org/@anthropic-ai%2fclaude-code