Claude Design — Anthropic launches design tool plus Opus 4.7, Routines, Advisor, and Managed Agents (Anthropic Labs)

TL;DR

Anthropic’s product channel announces five bundled launches keyed off Claude Opus 4.7: (1) Claude Design by Anthropic Labs — a conversational tool for prototypes, slides, and one-pagers with Canva / PDF / PPTX export and a Claude-Code handoff path; (2) the Opus 4.7 model itself (already filed separately); (3) Routines — scheduled, API-triggered, or webhook-triggered Claude runs with their own endpoints, hosted on Anthropic infra; (4) an Advisor API primitive that lets a Sonnet/Haiku executor consult Opus mid-run inside a single Messages API call; (5) Claude Managed Agents — a hosted harness with infrastructure, tools, and guardrails for production agent deployment, in public beta. The most concrete technical disclosure is the Advisor eval: Sonnet + Opus advisor scores +2.7 points over Sonnet alone on SWE-bench Multilingual at 11.9% lower cost per task.

Key claims

Claude Design is a conversational design surface that builds prototypes, slides, and one-pagers, refines through inline comments / direct edits / custom sliders, and exports to Canva, PDF, or PPTX or hands off to Claude Code. [tweet body]
Claude Design “reads your codebase and design files to build your team’s design system, then applies it automatically” — positioned as design-system-aware, not just per-asset generation. [tweet body]
Opus 4.7 is the vision model behind Claude Design; the launch tweet attributes the design capability specifically to Opus 4.7’s higher input resolution (>3× prior Claude) and better interface/slide/doc generation. [tweet body — see Introducing Claude Opus 4.7 for the underlying release post]
A new xhigh effort level lands on the API between high and max, with Task budgets (beta) added to help Claude prioritize work and manage costs across longer agentic runs. [tweet body]
Routines run on Anthropic’s web infrastructure (not the user’s machine), each get their own API endpoint, and can be triggered on a schedule, from an API call, or via webhook subscription to events such as GitHub. The advertised invocation shape is POST a message → session URL. [tweet body]
The Advisor strategy is a Messages-API-level primitive: a Sonnet or Haiku executor adds an “advisor tool” to its call; when it hits a hard decision mid-run it consults Opus, gets a plan, and continues — all within a single API request. [tweet body]
Headline Advisor eval: Sonnet + Opus advisor scores +2.7 percentage points over Sonnet alone on SWE-bench Multilingual while costing 11.9% less per task. [tweet body]
Claude Managed Agents is positioned as agent-harness-plus-production-infrastructure: define tasks, tools, and guardrails; Anthropic runs the agent on their infra. Public beta on the Claude Platform. [tweet body]
Early customer named: Notion. Teams delegate work to Claude inside the Notion workspace; “dozens of tasks run in parallel” and whole teams collaborate on the outputs. Private alpha. [tweet body]

Method

Product announcement tweet from @claudeai; no system card, no benchmarks beyond the one Advisor datapoint, no API schema disclosed. What is concretely described:

Claude Design: front-end built on Opus 4.7’s vision capability; flows are (a) describe → first version, (b) refine through chat / inline comments / direct edits / sliders, (c) export Canva/PDF/PPTX or hand off to Claude Code. Reads codebase + design files to derive a team design system.
Routines: declarative configuration (prompt + repo + connectors) once, then triggered via schedule, API call, or webhook. Each routine exposes its own API endpoint, so alerts / deploy hooks / internal tools can POST to Claude directly and get a session URL back. Webhook routines subscribe to event streams (GitHub explicitly named).
Advisor: added as a tool inside the Messages API; the executor model (Sonnet or Haiku) decides when to invoke it, the advisor (Opus) returns a plan, the executor resumes. Single API call boundary — no separate orchestration code on the user side.
Managed Agents: hosted agent harness + production infra, public beta. Configuration is tasks + tools + guardrails. No disclosed scaling limits, billing model, or eval numbers in the announcement.

Results

The only quantitative result in the tweet is the Advisor SWE-bench Multilingual datapoint: Sonnet + Opus-advisor configuration yields +2.7 percentage points over Sonnet-alone, at 11.9% lower cost per task. No baseline numbers are given (i.e., Sonnet’s absolute score and the dollar cost are not stated), so the +2.7 / −11.9% are the only quantities to cite.

Otherwise: Claude Design is in research preview on Pro/Max/Team/Enterprise. Routines and Managed Agents are in public beta on the Claude Platform. Notion Managed Agents is in private alpha.

Why it’s interesting

The Advisor primitive is the part worth tracking against the wiki: it formalizes the “small executor + large planner consulted on hard steps” pattern as a single-request API primitive rather than a user-side orchestration script. That’s the same axis as the K2.5 PARL learned-parallel-orchestrator and the RLM scaffold-side scaling already on Inference-Time Scaling — but instantiated at the closed-frontier deployment layer, where the orchestrator is not learned end-to-end, just exposed as a tool. The +2.7 SWE-bench Multilingual / −11.9% cost number is the first filed datapoint where a closed-source vendor monetizes the planner-as-tool composition pattern. Claude Design itself is a product-side analog of the design-tool cluster the team is watching (PRX Part 3 — Training a Text-to-Image Model in 24h, Introducing Recraft V4: Design Taste Meets Image Generation, PSDesigner: Automated Graphic Design with a Human-Like Creative Workflow) — but driven by a general-purpose VLM (Opus 4.7) rather than a specialist design model, with the differentiator being design-system-awareness (reads codebase + brand files) rather than pixel-level fidelity. Notable that the announcement explicitly attributes the design capability to Opus 4.7’s vision-input resolution lift (>3× prior Claude), the same axis that drove XBOW’s autonomous-pen-testing benchmark from 54.5 → 98.5% in the underlying release post (Introducing Claude Opus 4.7).