If you need AI agents to do a lot of ongoing tasks for your business, Anthropic has a new answer for you. The Claude maker has introduced Managed Agents, a service to help organizations create and deploy cloud-hosted knowledge work automations.<\/p>\n
Agents, for those who haven’t been following along, consist of machine learning models given access to software tools in an iterative loop. Claude Code is a coding agent that can emit programming code with the assistance of models like Opus 4.6 and permitted command line tools like bash, yoked together through a client-side harness \u2013 an orchestration tool.<\/p>\n
Those using Claude Code can create sub-agents that specialize in certain tasks, like frontend design. These are defined by Markdown files and YAML data \u2013 words that steer the underlying model toward training data related to functional interface patterns as opposed to code efficiency or some other goal.<\/p>\n
“An agent is a reusable, versioned configuration that defines persona and capabilities,” Anthropic explains<\/a> in its documentation. “It bundles the model, system prompt, tools, MCP servers, and skills that shape how Claude behaves during a session.”<\/p>\n Running an agent involves some degree of planning and configuration followed by monitoring and feedback \u2013 you give it a task, the agent attempts to comply, and it either asks further questions or proceeds to generate its interpretation of the desired response \u2013 until your token quota or your API budget has been exhausted.<\/p>\n Hence the appeal of Managed Agents: Anthropic is offering to make the agentic process a bit more hands-off and more scalable, which for organizations might have some appeal.\u00a0<\/p>\n “Shipping a production agent requires sandboxed code execution, checkpointing, credential management, scoped permissions, and end-to-end tracing,” the company said in a blog post<\/a>.<\/p>\n “Managed Agents handles the complexity. You define your agent’s tasks, tools, and guardrails and we run it on our infrastructure. A built-in orchestration harness decides when to call tools, how to manage context, and how to recover from errors.”<\/p>\n Where personal agent usage (for coding at least) tends to be semi-autonomous \u2013 you give the agent some tasks and check in as the model implements specific features \u2013 Claude Managed Agents is intended for longer periods of unsupervised action (a.k.a. spending).\u00a0<\/p>\n Managed agents are designed to muck around in their managed environment, reading files, running commands, browsing the web, and executing code without much oversight. The mundane aspects of LLM interaction \u2013 compacting sessions to free up context space, for example \u2013 are left to the machines.<\/p>\n Anthropic recommends Managed Agents for tasks that require a long time to complete and lots of tool calls, can operate in cloud-hosted secure containers, and benefit from persistent file and conversation data.\u00a0<\/p>\n The Managed Agents service isn’t just for coding, which remains the primary commercial use case for Claude to date. Anthropic suggests that its hosted ghost workers can handle a broad set of office tasks \u2013 a position underscored by Claude Cowork’s declaration of general availability<\/a> on Thursday.\u00a0<\/p>\n The AI biz emphasizes the general utility of its toilbots in a YouTube testimonial<\/a> from Notion product manager Eric Liu that describes how Notion uses Managed Agents to ship code and produce websites and presentations. This involves, for example, asking the managed agents to consolidate project assets, create Slack channels, research competitor home pages, and send emails with project timelines.<\/p>\n