Since April 2026, Anthropic launched three services – Routines, Agent Skills, and Claude Managed Agents – where AI agents are built and hosted directly with them. Yesterday at a San Francisco conference, they added multi-agent orchestration, outcomes, and dreaming. For Polish SMEs, this means one thing: for the first time, infrastructure stops being an excuse.
A quiet pivot the market hasn't named yet
For two years, everyone talked about models. Meanwhile, Anthropic has been systematically dismantling another problem: how to run an AI agent without your own stack. Claude Managed Agents (April 8), Routines in Claude Code (April 14), and yesterday three more mechanisms – these aren't cosmetic updates. This is a business-model shift: Anthropic stops selling the model alone and starts selling a complete service – sandbox, scheduler, key vault, orchestration. Four days ago they added a $1.5B joint venture with Goldman Sachs and Blackstone – to deploy Claude in mid-market companies "in days, not months."
What this concretely changes
A Polish entrepreneur from a 30-person company has been told until now that automating work with AI requires Azure, a DevOps engineer, and someone to write code. In Anthropic's model, each of these elements comes ready-made. Routines runs repetitive tasks in Anthropic's cloud – even with your laptop turned off. Agent Skills lets you upload your own procedures as folders of text files – no programming. Managed Agents hosts the entire agent, billing only the active seconds it works.
What was added yesterday
Three mechanisms worth knowing in English, because that's how they appear in the docs. Multi-agent orchestration – one conductor-agent splits a task across up to 20 specialized agents working in parallel. Outcomes – you write a rubric describing what success looks like, and a separate Claude instance verifies the result and tells the agent to fix it if criteria aren't met. Dreaming – the agent reviews its earlier sessions overnight, extracts patterns, and updates its own memory. The business consequence: the agent learns from its mistakes without your involvement.
What this pivot does not eliminate
It doesn't eliminate the need to know which process you want to hand to the agent. From my observations – confirmed by RAND and MIT reports – 80% of AI deployments fail to deliver business value, and 95% of companies using generative AI report no measurable return. That's not a technology problem, it's a lack of defined process. No Anthropic service changes that. Real risks also remain – vendor lock-in, data privacy, and AI Act requirements.
What's worth doing about it
Don't ask: "Can we afford the infrastructure for AI?" Ask instead: "Which of our processes is described well enough to hand to an agent today – and who's accountable when it makes a mistake?" Because the barrier Polish SMEs have been using as an excuse for the past two years just disappeared. The second one remains – and it was never technological. It was managerial.
Dear Reader. If you believe the topic above applies to your company and you'd like to talk with me in the Board's presence about how to sensibly use the new AI services in your company's reality, get in touch.
Leszek Giza