Agents moving into production

Artificial intelligence agents are starting to look less like chatbots and more like junior employees with logins, files, and a to-do list. On April 9, Anthropic moved two products in that direction at once: Claude Managed Agents for companies building autonomous workflows, and Claude Cowork on macOS with new enterprise controls. (anthropic.com) (9to5mac.com) A chatbot waits for a prompt and answers once. An agent is built to keep going across many steps, like opening files, checking tools, and finishing a task that can take minutes or hours instead of one reply. (anthropic.com 1) (anthropic.com 2) That sounds simple until a company tries to run one at scale. The hard part is not just the model’s intelligence, but the plumbing around it: permissions, tool connections, monitoring, retries, and guardrails when the agent gets stuck. (anthropic.com) Anthropic’s pitch is that companies should not have to build all of that plumbing themselves. Claude Managed Agents is a hosted service on the Claude platform that handles orchestration, sandboxing, and governance so a business can focus on the workflow instead of the infrastructure. (anthropic.com) (siliconangle.com) Anthropic says early users include Notion and Rakuten, which is a clue about the target customer. These are companies with enough internal process complexity that shaving weeks or months off agent deployment is worth paying for. (9to5mac.com) (the-decoder.com) Claude Cowork is the other half of the story because it puts the same idea on an employee’s desktop. Anthropic describes it as software for multi-step knowledge work on macOS, including research synthesis, document preparation, and file management, rather than a chat box in a browser tab. (anthropic.com) On April 9, Anthropic removed the “research preview” label from Claude Cowork and added enterprise features such as team administration and policy controls. That is the moment a lab demo starts turning into something an information technology department can approve. (9to5mac.com) The clearest sign that agents are moving into production is not the platform launch. It is that companies are now attaching revenue numbers to specific agent work. (qsrweb.com) (marketwatch.com) Deliverect, which says it serves more than 95,000 restaurant locations across 78 countries, launched a set of restaurant agents that rewrite menus, tune promotions, and watch digital ordering systems for problems. This week it pointed to a KFC campaign as proof that the model is already working outside a software lab. (qsrweb.com) (marketwatch.com) In Deliverect’s account, a marketing agent designed, launched, and optimized a KFC promotion with no human intervention and produced a 118% sales increase. Deliverect’s earlier December release said the pilot was for KFC Netherlands’ “Secret Box Meal” campaign and that the system automatically generated and activated rewards inside KFC’s ordering app. (qsrmagazine.com) (deliverect.com) That is the shift happening now. Last year, most agent demos were about what the software might be able to do; this week’s announcements are about who runs the infrastructure, who signs the enterprise contract, and whether the agent can move a sales number on a real dashboard. (anthropic.com) (qsrweb.com)