If your business needs an AI agent to perform the many ongoing tasks, Anthropic provides a new answer. The creators of Claude have introduced Managed Agent, a service that helps organizations create and deploy cloud-hosted knowledge work automation.
For those who don’t already know, an agent consists of a machine learning model that is given access to software tools in an iterative loop. Claude Code is a coding agent that can output programming code with the help of models such as Opus 4.6 and authorized command line tools such as bash, coupled through client-side harnesses (orchestration tools).
Claude Code allows users to create specialized subagents for specific tasks, such as front-end design. These are defined by Markdown files and YAML data. These words orient the underlying model to training data related to functional interface patterns, rather than code efficiency or other goals.
“An agent is a reusable, versioned configuration that defines personas and capabilities,” Anthropic explains in its documentation. “It comes bundled with models, system prompts, tools, an MCP server, and skills that shape Claude’s behavior during a session.”
Running an agent involves some planning and configuration, followed by monitoring and feedback. When you give an agent a task, it tries to follow through and continues asking more questions or generating interpretations of the desired response until its token allocation or API budget is exhausted.
So the appeal of a managed agent is: Anthropic suggests making the agent process a little more manual and more scalable. This may be attractive to organizations.
“Shipping a production agent will require sandboxed code execution, checkpointing, credential management, scoped permissions, and end-to-end tracing,” the company said in a blog post.
“Managed agents handle complexity. We define tasks, tools, and guardrails for agents, and we run them on our infrastructure. A built-in orchestration harness determines when to invoke tools, how to manage context, and how to recover from errors.”
Personal agent usage (at least for coding) tends to be semi-autonomous. Give the agent some tasks and check in when the model implements a certain functionality. Claude managed agents are intended for long-term, unmonitored actions (aka spending).
Managed agents are designed to read files, run commands, browse the web, and execute code within a managed environment without much supervision. Day-to-day aspects of LLM interactions are left to the machine, such as session compaction to free up context space.
At Anthropic, we recommend managed agents for tasks that take a long time to complete and require many tool calls. It can run in secure containers hosted in the cloud and benefits from persistent file and conversation data.
Managed agent services aren’t just for coding. Coding continues to be Claude’s main commercial use case so far. Anthropic suggests that hosted ghost workers can handle a wide range of office tasks. This position is underscored by Claude Cowork’s general availability declaration on Thursday.
The AI industry highlights Toilbot’s general utility in a YouTube blurb by Notion product manager Eric Liu, which explains how Notion uses managed agents to distribute code and create websites and presentations. This includes, for example, asking managed agents to consolidate project assets, create a Slack channel, research a competitor’s homepage, or send an email with a project timeline.
You get all this for the low price of standard platform fees, plus an additional $0.08 per active runtime session hour. ®
