Managed Agents
Managed Agents are AI programs that execute tasks autonomously while being coordinated by an overarching system. They are a central concept in modern AI applications, in which multiple specialized programs work together through a division of labor.
A Managed Agent is an AI program that handles a task independently — meaning it doesn’t just answer a question, but plans steps, uses tools, and makes decisions. The crucial addition lies in the word “Managed”: an overarching control system, often called an orchestrator, assigns the agent its task, monitors it, and coordinates it with other agents. A Managed Agent therefore never works entirely alone, but is embedded within a larger system made up of several such programs. The individual agents are usually specialized for particular tasks, similar to departments within a company.
Division of labor as a strength
A single AI model quickly reaches its limits when a task demands many different capabilities. A Managed Agent system solves this problem through specialization. One agent researches on the internet, another writes code, a third checks the result — and the orchestrator holds everything together.
This is more efficient than packing everything into a single model. It also makes the system easier to maintain: if an agent makes mistakes, it can be swapped out without having to touch the rest. For companies, this also means they can specifically improve or replace individual agents as soon as better models become available.
A common misconception is to assume that “Managed” means a human is steering the agent in real time. Often the opposite is true: the management is handled by an automated system. The human only sets the goal at the start — what happens afterward runs autonomously.
Structure of a Managed Agent system
At the center is the orchestrator. It receives a task, breaks it down into subtasks, and assigns them to the appropriate agents. Each agent has access to certain tools — these can include web searches, code execution, database queries, or sending emails. The agent itself chooses which tool to use and when.
Communication between the orchestrator and the agents usually takes place via structured messages. An agent reports its result back, the orchestrator evaluates it and decides whether the task is complete or whether a further step is needed. This feedback loop is crucial: if a step fails, the system can try again or assign the task to a different agent.
An important difference from simpler approaches: in the classic prompt-and-response principle, the model responds once to an input. Managed Agents, on the other hand, carry out many such steps in sequence, with each step building on the result of the previous one. This enables significantly more complex tasks — for example, a complete market analysis including data retrieval, evaluation, and report generation.
Managed Agents in products and headlines
Managed Agents are increasingly appearing in concrete software products. Microsoft's Copilot Studio allows companies to build their own agents and have them managed by a central system. Salesforce offers a similar platform for sales and customer service with “Agentforce.” The underlying principle is the same in both cases: specialized agents under central control.
In research, frameworks such as LangGraph or AutoGen are well known, allowing developers to assemble their own Managed Agent systems. There, agent networks can be defined with just a few dozen lines of code, which makes the technology accessible even outside large corporations.
In business news, the term often appears in connection with the automation of office work. Analysts estimate that Managed Agent systems could, in the medium term, take over entire workflows that are today still handled entirely by humans — from invoice processing to software development. How reliably and safely such systems act in doing so is one of the central open questions.