This Week in Coding Agents: September 21, 2026

Week 39, 2026 · Published September 21, 2026

Claude Code	2.1.278
Codex CLI	0.155.1
Gemini CLI	0.60.0

The short version

Claude Code’s changelog lists 2.1.278 as the top release (September 19). npm latest is 2.1.278. 2.1.277 reads AGENTS.md when the project has no CLAUDE.md. 2.1.278 defaults auto mode to the server-side classifier, which does not charge for classifier overhead.
Codex CLI 0.155.1 (September 18, tag rust-v0.155.1) is the current stable. 0.155.0 (September 17) adds experimental /voice through /experimental. New local TUI sessions now leave reasoning summaries off by default. The releases list shows 0.156.0-alpha.14 as a prerelease.
Cursor has no version number and no new changelog entry this week. The top item is still Cursor Projects from September 10.
GitHub Copilot has no version number. Auto model selection now has efficiency, balance, and intelligence tiers. Six models drop on October 19, 2026. Code review, budget increase requests, and several admin metrics pages shipped.
Gemini CLI 0.60.0 shipped on September 15. The notes I read are path, sandbox, and MCP OAuth hardening. No new model name appears on that page.

Claude Code

2.1.271 (September 14) through 2.1.278 (September 19). I opened the live CHANGELOG.md and the memory docs. The first heading is 2.1.278. npm latest is 2.1.278.

2.1.278 (September 19). Auto mode for Claude API and Enterprise users, and on Bedrock, Vertex, Foundry, and gateways, now defaults to the server-side classifier. The billing page says those server checks are not billed when the server performs them. CLAUDE_CODE_AUTO_MODE_SERVER=0 opts out on Bedrock, Vertex, Foundry, and gateways. If the server cannot reach the session, Claude Code falls back to billed local classifier requests and shows a notice before the first checked action. /status adds an Auto mode server row (Enabled or Disabled). Pro, Max, and Team plans never see the notice.

2.1.277 (September 18). AGENTS.md support: in a project with no CLAUDE.md, Claude Code reads AGENTS.md instead. Change the default under “Project instructions” in /config. The changelog marks this as not yet on Bedrock, Vertex, or Foundry. The memory docs I opened match that and add more: CLAUDE.local.md and .claude/CLAUDE.md also count as a CLAUDE.md, so they block the fallback; ~/.claude/CLAUDE.md, managed CLAUDE.md, and .claude/rules/ do not. Default value is claude-md-or-agents-md. Other values: claude-md-and-agents-md, claude-md, managed-only. Direct AGENTS.md load needs 2.1.277 or later. It is also off in the first session after upgrade, when telemetry or feature flags do not reach Anthropic, and when the built-in agents-md plugin is disabled. A CLAUDE.md that already has @AGENTS.md can stay; the docs say Claude will not read the file twice. The deprecated TaskOutput tool is gone. Claude reads a background task’s output file with Read. taskOutputMaxChars and TASK_MAX_OUTPUT_LENGTH no longer have any effect. Fable always appears in /model on the Anthropic API and is greyed out only when the org disables it.

2.1.276 (September 18). One note: every request failing with 400 … Input tag 'advisor_20260301' when ANTHROPIC_BASE_URL points at a proxy or gateway. That was a 2.1.275 regression.

2.1.275 (September 17). A send-now key (ctrl+enter, or ctrl+x ctrl+s) interrupts the current turn and sends every queued message at once. Skills and plugins enabled on your claude.ai account sync into terminal sessions signed in with that account. Opt out with syncClaudeAiSkills: false or syncClaudeAiPlugins: false. /plugin install <plugin> --marketplace <source> offers to add the marketplace before it installs. /update-config now writes Edit(path) permission rules instead of Write(path) rules that file checks did not match. Plugins from npm are fetched with npm pack --ignore-scripts and integrity-verified.

2.1.274 (September 16). CLAUDE_CODE_MCP_STARTUP_WAIT_MS bounds how long the first non-interactive turn waits for connecting MCP servers (0 means do not wait). New OTel: effort on the claude_code.llm_request span, and claude_code.managed_settings_resolved (redacted settings and digests with OTEL_LOG_MANAGED_SETTINGS=1). Streamable HTTP MCP tool calls honor a longer per-server timeout instead of dying at about 5 minutes. MCP servers configured as http that only speak legacy HTTP+SSE connect when the first answer is 422 or another 4xx. Bedrock, Vertex, Foundry, and telemetry-disabled installs now use the v2 MCP client and MCP 2026-07-28 negotiation with direct HTTP servers by default (MCP_SDK_GENERATION=v1 or MCP_PROTOCOL_NEGOTIATION=legacy to opt out). "type": "sdk" MCP entries in .mcp.json are skipped with a warning.

2.1.273 (September 15). Opt-in gateway hint headers with CLAUDE_CODE_GATEWAY_HINT_HEADERS=1: x-claude-code-request-class, x-claude-code-agent-type, x-claude-code-prev-tool-durations, x-claude-code-compaction, x-claude-code-context-compacted. Fork a claude --remote-control or /remote-control session from the Claude app; the fork runs as a background session on your computer. Auto mode on Bedrock, Vertex, and Foundry used the local classifier by default in this build (CLAUDE_CODE_AUTO_MODE_SERVER=1 for the platform classifier). 2.1.278 flipped that default. A 2.1.268 change that checked Read and Edit deny rules on Bash lines the permission checker cannot analyze (eval, env -C) is reverted. Commands like time -p make build prompt again instead of being denied. OTEL_LOG_TOOL_DETAILS=1 now includes real agent, skill, plugin, and MCP server names on cost and token metrics.

2.1.272 (September 14). The only note on the page: bug fixes and reliability improvements.

2.1.271 (September 14). Fast mode in Remote sessions (cloud and self-hosted runners): the host’s setting or /fast typed in the session applies where the org allows it. omitClaudeMd in agent frontmatter and --agents JSON lets custom and plugin subagents skip user, project, and local CLAUDE.md files. Managed policy files still load. Per-command allowed_domains on Bash, PowerShell, and Monitor in auto mode with sandboxing: the hosts that command needs are reviewed with it and opened for it alone. claude plugin install and claude plugin update accept --accept-command <sha256> for the exact command a previous --json run displayed, instead of -y. modelPricing and the Claude apps gateway pricing block accept a multiplier above 1, up to 10.

Codex CLI

0.155.1 (September 18, tag rust-v0.155.1) and 0.155.0 (September 17, tag rust-v0.155.0). I opened both release pages. The releases list shows 0.156.0-alpha.14 as a prerelease. src/data/tool-versions.json records latest as 0.155.1.

0.155.1 (September 18). New local TUI sessions leave reasoning summaries disabled by default. That fixes request rejection by providers that do not support them. Explicit reasoning-summary settings stay respected.

0.155.0 (September 17). Experimental /voice conversations with live transcripts and microphone controls on supported builds, enabled through /experimental. The TUI shows live reasoning summaries in the status row and completion timestamps after successful turns. The agents overview can hide, archive, and delete tasks, and it shows worktree ownership plus confirmed deletion of clean managed worktrees. Local TUI sessions on supported Macs can use Touch ID for MCP request verification. Daemon updates are configurable, and codex app-server daemon update is a first-class command. Saved threads and active goals can recover after a daemon restart. Amazon Bedrock can obtain AWS credentials from configured commands, with caching, expiration-based refresh, and authentication recovery.

GitHub Copilot

No version number. I opened each changelog URL below.

Copilot code review: An improved review experience (September 18). Generally available. The overview comment now groups findings as Open, Resolved since last review, and Previously missed. Comments get titles. Auto-resolution honors a reply that asks to leave an issue open, and it records Won't Fix or Incorrect from later commits. Committing an eligible batch of suggestions (including a mix with non-Copilot comments) gets a generated title and optional description.

Weekly releases, September 14 (September 18). Recap of the auto-selection tiers, code review updates, and VS Code Agents metrics. Copilot app: a Sentry canvas that takes you from a crash report to a code fix and PR prep. VS Code 1.138: local Dev Containers in the Agents window (rolling out; needs Docker and a supported config); opt-in automatic cleanup of inactive sessions after their PRs merge (preview); create a pull request from an Agent Host session without leaving the Agents window.

Upcoming deprecation of selected models in mid-October (September 18). Drops on October 19, 2026 across Chat, inline edits, ask and agent modes, and completions: Gemini 3.7 Flash (use Gemini 3.8 Flash), GPT-5.5 and GPT-5.4 (use GPT-5.6 Sol), GPT-5.4 mini and GPT-5 mini (use GPT-5.6 Luna), Grok 4.5 (use Grok 4.6). Under default model enablement, the replacements are on for Copilot Enterprise and Business unless an admin turned off the global default or disabled the model. The page says you do not have to remove the old ones.

Copilot impact dashboard now shows feature engagement (September 17). Enterprise dashboard plus 28-day aggregate report fields: copilot_feature_engagement, totals_by_feature (code completion, agent edit, passive and active code review, cloud agent, CLI, app), and users_in_phase_28d. A user counts as engaged with a feature after at least two days in the 28-day window. Not added to user-level reports. Access: enterprise owners, billing managers, organization owners, and custom roles with View Copilot Metrics, with the Copilot usage metrics policy enabled.

Agentic CLI customizations now in the usage metrics API (September 17). Enterprise and organization 1-day and 28-day reports add totals_by_skill, totals_by_custom_agent, totals_by_mcp, totals_by_slash_cmd, totals_by_plugin (top five each, with interaction_count) and matching distinct_*_use_count fields. GitHub-provided names are shown. Customer-defined skills, agents, MCP servers, and plugins are grouped as other; slash commands stay under custom. MCP interaction_count is connection or reconnect attempts, not tool calls. Plugin totals are a subset of skill totals. Do not add the two.

Copilot budget increase requests are generally available (September 16). Copilot Business and Enterprise on usage-based billing. Not for enterprises with managed users. A member who hits the AI credit limit can request more. Owners and billing managers approve, adjust, or deny under “Requests from members.” Approval sets the new amount and restores credits.

GitHub Copilot suggests custom properties definitions (September 15). Public preview for Copilot Business and Enterprise. Creating a custom property at the enterprise or organization level can suggest allowed values (internet-facing to yes / no, FedRAMP to compliance values). Owners can turn this off with the Repository custom property suggestions Copilot policy.

Configure cost and quality in Copilot auto model selection (September 14). Three tiers, rolling out in VS Code, Copilot CLI, and the Copilot app: efficiency (cost and speed), balance (cost, quality, and latency), intelligence (quality for harder tasks). All three pick from the same model set, per prompt. Usage is billed on the model auto selects. Paid subscribers still get a 10% discount on usage billed through auto.

Gemini CLI

0.60.0 (September 15). I opened the v0.60.0 release. The releases list shows that tag as the current stable release. Nightly tags and v0.61.0-preview.0 sit above it.

Notes I saw: web-fetch destination checks, RFC 9207 issuer identification in MCP OAuth, macOS Seatbelt temp-dir isolation, sandbox settings-dir isolation, extension path and env consent hardening, workspace and symlink boundary checks, system-config ownership checks, NTFS 8.3 short-name mitigation, and envelope metadata on untrusted tool output. No new model name appears on that page.

What this means for your setup

One AGENTS.md, every tool still says Claude Code does not read AGENTS.md on its own. After 2.1.277 that is false for a repo that has AGENTS.md and no CLAUDE.md / CLAUDE.local.md / .claude/CLAUDE.md, except Bedrock, Vertex, Foundry, telemetry-off sessions, and the first session after upgrade. Refresh that page.
New repo agent setup still treats a one-line CLAUDE.md with @AGENTS.md as required. Keep the import if you already have CLAUDE.md or you run on Bedrock, Vertex, or Foundry. You can skip writing CLAUDE.md on 2.1.277+ when AGENTS.md is the only project file. Refresh the Claude Code section.
Instruction files is still a draft that only describes the bridge. Context is a budget still lists CLAUDE.md and CLAUDE.local.md only. Both need a refresh for the 2.1.277 fallback.
If you add a CLAUDE.local.md next to a repo that relies on AGENTS.md, Claude Code stops reading AGENTS.md unless you set Project instructions to claude-md-and-agents-md in /config.
If you still set taskOutputMaxChars or TASK_MAX_OUTPUT_LENGTH, 2.1.277 ignores both. No learn or playbook page here names those knobs.
Auto mode users on 2.1.278: run /status and read the Auto mode server row. If a gateway strips unknown fields, you will see the billed-fallback notice. Set CLAUDE_CODE_AUTO_MODE_SERVER=0 only when you already know the gateway cannot pass server checks.
Codex users who want spoken turns: /experimental, then /voice. 0.155.1 turns reasoning summaries off for new local TUI sessions. If you relied on them, set the reasoning-summary option yourself.
Copilot users still on Gemini 3.7 Flash, GPT-5.5, GPT-5.4, GPT-5.4 mini, GPT-5 mini, or Grok 4.5 should switch before October 19. Admins who disabled the global default need to enable the replacement models in policy.
Copilot auto users: pick efficiency, balance, or intelligence. The page says a simple prompt can still land on a small model even on the intelligence tier.
Permissions and unattended runs and Unattended agents in CI still match the --permission-mode dontAsk, --permission-prompts none, --bare, and codex exec --sandbox workspace-write --ask-for-approval never flags I opened this week. 2.1.273 undid the 2.1.268 deny of unanalyzed Bash lines. Those pages do not document that case.

Pages we updated this week

None this week.