Claude Code Finally Ships a “Concise” Output Style
Claude Code v2.1.237 landed this morning with a built-in Concise output style. The description is refreshingly blunt: Claude “leads with results and skips preamble and narration, while doing the work just as thoroughly.” You select it under Output style in /config. No custom style file, no CLAUDE.md incantation, no begging in the system prompt.
This is a small release with a long backstory. “Stop telling me what you are about to do” has been one of the most persistent complaints about agentic coding tools, and until now the workaround was a hand-rolled output style that people copied off each other and re-tuned every few versions. Making it a first-party toggle matters mostly because it is now a supported behavior rather than a prompt hack that a model update can quietly break. The other item in 2.1.237 is narrower but worth knowing if it bites you: prompt caching is fixed for sessions running through an LLM gateway or a custom base URL. If you route Claude Code through a proxy and your cache-hit rate has looked wrong, that was the bug.
Files, Skills and Admin User Management All Leave Beta
The Developer Platform cleared three betas in a single release note. The Files API is now GA — /v1/files and Messages API requests referencing an uploaded file no longer need the files-api-2025-04-14 header. The GA response format brings real lifecycle controls: file expiration via expires_in_seconds with expires_at on the object, page/next_page pagination, and an ids[] filter on list. Storage is 1 TB per organization at 500 requests per minute.
Agent Skills and the Skills API (/v1/skills) also went GA, dropping the skills-2025-10-02 header including on Messages requests that load skills through container. And Admin API user management for Claude Enterprise — members, invites, groups and custom roles — is GA, retiring ce-user-management-2026-07-13. In all three cases requests that still send the old header keep working unchanged, so nothing breaks today. But if you have been holding a skills or files integration back because it sat behind a beta flag, that objection is gone. The Files API in particular now has the expiry and pagination semantics you need to run it as infrastructure rather than a demo.
Domain Allowlists, Sandbox Memory, and a Rebuilt Session Viewer
Three Managed Agents changes shipped alongside the GA wave, and the first one closes a real gap. You can now restrict which sites an agent’s web_search and web_fetch tools can reach, by setting allowed_domains or blocked_domains on the tool’s entry in the agent_toolset_20260401 configs array. web_fetch also picks up max_content_tokens and web_search accepts user_location. Existing entries that pass only name, enabled and permission_policy keep working.
Second: sessions running in a self-hosted sandbox can now attach memory stores. The Python, TypeScript and Go SDK workers download each attached store into the sandbox at its mount_path and sync the agent’s changes back — which is what makes long-running self-hosted agents stateful without you building the persistence layer yourself. Third, the Console session viewer has been redesigned: a timeline minimap, a transcript grouped by model request, and an Inspector panel covering session cost, raw events, per-tool statistics, mounted resources and per-thread activity. That last one is the difference between debugging an agent and guessing at one.
A Default Model You Can Actually Set, and Sessions That Ping You When They Finish
v2.1.236 was the big one this week, and two additions stand out. ANTHROPIC_DEFAULT_MODEL sets the model new sessions start on while leaving a /model pick free to override it and persist across restarts — the distinction from ANTHROPIC_MODEL, which pinned you. If you have been fighting your own env vars to get a sane team default, this is the fix.
Second: notify_when_idle on cross-session SendMessage. You can ask another Claude Code session on the same machine to send one notice when it next goes idle — opt-in, one-shot, no polling, macOS and Linux. Combined with last week’s @-mention-another-session feature, parallel sessions are quietly turning into a coordination primitive. Also in 2.1.236, and worth reading if you rely on the sandbox: on macOS, wildcard read-deny rules like **/.env now take precedence inside allowed read regions, cover matched directories’ contents, and cannot be bypassed by renaming the denied file. There are also fixes for the /model picker rendering taller than the terminal, fullscreen renderer failures, and a runaway-recap bug now capped at 400 characters.
Auto Mode Gets Harder to Fool
Three changes to auto mode landed together and they all point the same direction. Monitor allow rules are now set aside while auto mode is active, so Monitor commands get reviewed on the same footing as Bash commands rather than sailing through on a pre-approval written for a different context. Second, auto mode on Bedrock, Vertex AI and Foundry — and when telemetry is disabled — now uses the same classifier defaults as the Claude API, including severity-scored classification. Third, the git status check can no longer be fooled by a repository’s status.showUntrackedFiles=no setting into reporting a clean tree.
None of these is a headline feature and all three are the kind of thing you only notice when they are missing. The pattern is the interesting bit: auto mode shipped with the permissive path being whatever the fastest deployment allowed, and it is now being tightened toward parity, with the cloud-provider deployments catching up to the first-party one. If you evaluated auto mode on Bedrock or Vertex earlier and found the safety behavior weaker than the docs implied, re-run that evaluation.
Cowork Is on Mobile for Every Paid Plan Now
Two days after the Chrome side-panel rollout, Cowork is available on mobile and web across all paid plans. The honest framing from Anthropic is that mobile is strongest for starting tasks, checking status, steering mid-task and picking up results — deep local-folder access and desktop-file workflows stay the desktop app’s territory. That is the right expectation to set. Nobody is building a spreadsheet from a phone.
The part that actually changes behavior is remote sessions, now in beta: files and session state save to your Claude account and sync across devices, tasks keep running in the background even with your device off, and when Claude hits a decision point that needs a human, it sends an approval prompt to your phone instead of stalling. That converts the most common failure mode of long agentic runs — finishing the reasoning and then sitting idle for four hours waiting for someone to come back to the laptop — into a push notification. Cowork went desktop-only to browser, mobile and background-remote in about eight days.
Eight Straight Days With an Incident
Yesterday’s degraded performance on Opus 5 and Haiku 4.5 was the tenth logged incident since August 12 — eight consecutive days without a clean one. The run includes the multi-service outage on the 16th that ran roughly 42 minutes from 21:58 to 22:40 UTC, and degradation on the 18th spanning claude.ai, platform.claude.com, the API, Claude Code and Cowork. Anthropic has not published a root cause for the cluster.
Context matters here: prior to this stretch, 2026 had a fairly ordinary incident profile — March 2, June 2, June 18, July 29. Ten in eight days is a step change, not noise, and it arrives in the same month the company is talking to public markets. The practical advice has not changed since yesterday and is worth repeating because the streak has not broken: if Claude sits in a production path, confirm you have a failover model configured and tested, and check that your retry logic backs off rather than amplifying an incident. Availability is the risk you are carrying this month.
The Watermark Is Live, and the Real Deadline Is December 2
Worth a status check now that the news cycle has moved on: every Claude product released from August 2 onward embeds machine-readable marking in generated content, applied to text and file output globally rather than only in the EU. The driver is the EU AI Act Transparency Code, which took effect August 2 and requires AI-generated or edited content to be markable by other systems. Anthropic chose to ship one behavior everywhere instead of geofencing it.
Two details that get lost in the summaries. First, the marking is designed to survive copy and paste — it travels with the text, not with the file wrapper. Second, and more consequentially for anyone building on older models: Anthropic has until December 2, 2026 to roll watermarking out to Claude models that predate August 2. That is the date to put in the calendar, because it is when the behavior becomes uniform across the fleet rather than split between marked and unmarked outputs depending on which model version answered.
Everything Left Beta at Once. That Is Not an Accident.
Files, Skills, the Skills API, Admin API user management — four surfaces that had been sitting behind beta headers for months to a year all cleared GA in the same release note. Add the Managed Agents domain allowlists, the sandbox memory stores and the rebuilt Console session viewer, and the shape of the week is clear. This is not a feature push. It is a platform tidying its edges.
Beta headers are a promise that the contract might change. Enterprises read them exactly that way, which is why “it is still in beta” is the single most common reason a Claude integration sits in a proof-of-concept folder instead of a production repo. Retiring three of them in one day, while keeping the old headers working so nothing breaks, removes that objection at essentially zero migration cost. The domain allowlists do the same job from the security side: the first question any review board asks about an agent with web access is “what can it reach,” and until yesterday the answer was “the internet.”
The timing is not subtle. A company two months from a listing, with a run rate that reportedly crossed $65 billion, has every reason to convert experimental surfaces into contractual ones before anyone diligences the platform. And there is a second read worth holding: the Concise output style arriving the same week is a small piece of the same instinct. Both are about reducing friction that engineers have been routing around by hand. The bet is that the growth from here comes less from new capability and more from removing the reasons people have not deployed the capability that already exists. That is a mature-product bet — and given a status page that has logged an incident every day for eight days running, it may also be the only kind of bet the infrastructure can currently support.