AI Apps
Microsoft Agent Framework
The framework behind a multi agent workload: planning, registered tools, and state that survives a long running task.
Product
Microsoft Agent FrameworkMicrosoft Agent Framework is the code path for work that a single retrieval step and a single prompt cannot finish. It provides planning, a registry of tools an agent may call, delegation between specialist agents, and state that survives a task running longer than a request. Where a grounded assistant answers from an index, an agent workload reads several systems, performs a sequence of steps, and has to be inspectable afterwards. Observability is part of the framework rather than added later, because a multi step run that produced the wrong answer is only debuggable if the steps were recorded.
How it works
- A request is received and a planner decides which specialist agents and which steps the task requires.
- Specialist agents hold a narrow remit each, so the prompt and the tools of one do not have to cover the whole domain.
- Tools are registered with a declared contract, and an agent may call only what has been registered to it.
- Long running work is carried on durable orchestration, so a task survives a restart rather than beginning again.
- Shared state and memory are held outside the agents, commonly in Azure Cosmos DB, so a step can read what an earlier step established.
- Every step, tool call, and intermediate result is traced, so a wrong outcome is inspected rather than reproduced by guesswork.
- The run is evaluated end to end and not only at the final answer, because a plausible answer reached through a wrong step is still a defect.
A problem it solves
The question crosses several systems and takes a sequence of steps, so a single prompt over a single index either answers part of it or answers confidently from the wrong place. Microsoft Agent Framework makes the sequence explicit: which agent handled which step, which tools were called, and what each step returned, which is also what makes the result reviewable.