Anthropic pushes managed agents and hybrid advisor strategy
Anthropic launched a managed infrastructure for autonomous agents and unveiled an ‘advisor strategy’ that pairs its Opus model with lighter models for cost-effective agent intelligence, listing early adopters like Notion and Rakuten. The move packages autonomy with hosting, controls and support—shifting the sales pitch from raw model performance to operational readiness for enterprises. (the-decoder.com) (x.com)
Anthropic is no longer just selling a smarter model. It just launched Claude Managed Agents, a public beta service that lets companies run autonomous agents on Anthropic’s own infrastructure instead of building the plumbing themselves. (anthropic.com) That plumbing is the messy part most demos skip. Anthropic says a managed agent bundles a session log, a harness that keeps calling the model and routing tools, and a sandbox where the agent can run code and edit files. (anthropic.com) The pitch is simple: companies write the agent logic, and Anthropic handles hosting, orchestration, and governance. The Decoder reports the service runs only on Anthropic’s infrastructure and adds $0.08 per session hour on top of normal token charges. (the-decoder.com) Anthropic also attached customer names to the launch. The company said Notion is using the system for workspace delegation, Rakuten is building enterprise agents in Slack, and Sentry is using it for automated debugging. (the-decoder.com) The second part of the launch is a pricing trick disguised as an architecture trick. Anthropic’s new advisor strategy lets a cheaper model do most of the work and call Claude Opus 4.6 only when it hits a hard decision. (buildfastwithai.com) Claude Opus 4.6 is Anthropic’s top-end model, and the company says it is built for long, multi-step work like coding, search, and finance. Anthropic lists pricing for Opus 4.6 at $5 per million input tokens and $25 per million output tokens, which is why companies do not want it supervising every tiny step. (anthropic.com) In the advisor setup, Sonnet 4.6 or Haiku 4.5 acts like the worker on the shop floor, and Opus acts like the senior partner who gets pulled in only for the hard calls. A third-party writeup describing the launch says Sonnet with an Opus advisor improved SWE-bench Multilingual from 72.1% to 74.8% while cutting cost per agentic task by 11.9%. (buildfastwithai.com) Anthropic has been moving toward this for months. When it introduced Claude 4 in May 2025, it framed Opus as the model for long-running agent workflows, and when it upgraded to Opus 4.6 on February 5, 2026, it added a 1 million token context window in beta for bigger, longer jobs. (anthropic.com 1) (anthropic.com 2) So the company’s sales pitch is changing. Instead of asking buyers to compare benchmark scores in a vacuum, Anthropic is packaging the model, the runtime, the guardrails, and the support into something that looks more like a cloud service than a chatbot. (wired.com) (unite.ai) That puts Anthropic in the part of the market where budgets get bigger and contracts get stickier. Once a company’s agents are running inside Anthropic-managed sessions, sandboxes, and monitoring loops, switching vendors stops being a model swap and starts looking like a migration project. (anthropic.com) (startupfortune.com)
