ClaudeDevs

Three updates to Claude Managed Agents: work done in a self-hosted sandbox can be saved to memory, web_search and web_fetch accept allowed_domains or blocked_domains, and the console session viewer is rebuilt for multi-agent runs with per-thread and per-session cost.

A new cookbook pairs Claude Managed Agents with AG-UI, an open agent to user interaction protocol. The adapter maps each chat thread to a managed session and streams text, tool calls and thinking through to your own interface.

Anthropic slipped an obviously dangerous command into a permission prompt for 1,053 paid testers. They caught it 13.6% of the time, dropping toward 5% after fifty prompts, while a separate classifier blocked the same commands 89% of the time and did not degrade as sessions ran long.

Claude Code ships /design as a research preview in both CLI and Desktop, bringing Claude Design's artboard workflow to the terminal on top of artifacts — generate UI options, edit the one you want, then have Claude build it. Live on Pro, Max, Team and Enterprise after updating.

Claude Code's screen reader mode adds labeled lines for you, Claude and tools, numbered menus answered with a digit, and a bell when input is needed, persistable via settings or an environment variable.

Claude Managed Agents can now have an advisor: one roster line gives a working agent a stronger model it can call mid-session for a second opinion.

On the Chartography benchmark of 100 questions over dense real-world charts, giving Fable 5 a zoom tool lifts accuracy from 29% to 73%, and Sonnet 5 from 13% to 44%.

Claude Managed Agents gains inference geography control: pin a session to US-region capacity at 1.1x the rate, or run globally wherever capacity exists at standard pricing.

Claude Managed Agents now load skills from attached repositories: anything already kept in .claude/skills/ for Claude Code is picked up automatically when a managed session starts.

Claude Managed Agents supports outcome-driven loops: hand Claude a rubric for task completion and it keeps working until the rubric is satisfied, enabling long-horizon autonomous work with self-correction.

Anthropic's Admin API now ships in the official SDKs and the `ant` CLI, covering members, workspaces, invites, API keys and org rate-limit reads. It authenticates with either an sk-ant-admin key or an org:admin OAuth token, with service-account and federation endpoints OAuth-only.

Anthropic is prototyping Function Hooks for Claude Code, an extension point intended to make plugins substantially more capable. Not shipped; the design is open for comment on GitHub issue #91870, with videos demonstrating the intended behavior.

Anthropic opened a directory submission portal for Claude plugins on paid plans. Submit a remote MCP connector or a GitHub plugin bundle, get auto-validation and a safety scan, track review, and see installs by product and version. MCP 2.0 extensions are supported.

Anthropic resumes billing for requests its safety classifiers block before a response, limited to biology, distillation-attack, and frontier-LLM-development categories with under 0.1% false-positive rates, citing recent coordinated attacks.

Anthropic's Claude Code team introduces mods: TypeScript extensions shipped in plugins that alter behavior, customize the UI and add features. Examples include a context-window forecast, a risky-command guard and a diff replay pane.

Anthropic opens claude.dev, a developer blog and video hub with playbooks, Claude Code and API guides, skills and agent engineering write-ups, including recent posts on Opus 5.5 task costs and automated eval design.

Anthropic breaks down the exact loop, per-thread benchmarks, steering, and safety guardrails, they used to have Claude drive a two-week performance sprint that shipped over 3,000 changes with zero rollbacks.

Anthropic's playbook on building trustworthy evals and improving against them: production-like task sampling, headroom and variance checks, grader validation, and overfitting guards. Worked examples include cutting support-ticket cost to about a fifth and lifting a skill eval from 66% to about 88%.

Anthropic's Managed Agents add dynamic workflows in public beta: a lead agent writes a multi-phase plan, runs it across many agents, then merges results. Enabled via a multiagent type, with a bug-hunting benchmark and a warning to start with scoped tasks.

A hands-on playbook for running Claude Code in cloud sessions: parallel tasks on isolated VMs, seven suited workflows, local versus cloud tradeoffs, environment setup so agents can verify work, and a GitHub connection checklist.

Anthropic moves computer use, the browser tool, the Skills API and the Files API to general availability, adding several actions per turn, element level browser control, file expiry and higher rate limits.

Anthropic takes Claude Code cloud sessions out of research preview, letting agents keep running on hosted infrastructure without a local machine, and gives Pro and Max subscribers a one-time $100 or $250 credit through October 7.

Claude Code's new `claude plugin eval` command lets developers draft test cases, run their plugin or skill against them, score the runs, then rerun without the plugin to see the delta, shown in-terminal and as an HTML report.

Claude Managed Agents adds a session viewer (`ant beta:sessions connect`, with a `--web` UI) and an `auto` mode that reviews each tool call against stated intent to decide whether to run, deny, or ask.