Does every specialist really need its own AI agent?
A Microsoft engineering example replaces some specialist agent loops with distributed skills and MCP tools, showing how teams can preserve domain services while simplifying where reasoning happens across a workflow.
Microsoft published an architecture comparison that moves selected specialist instructions into an orchestrating agent while leaving domain services behind MCP interfaces. The example compares this with delegating to autonomous specialists through A2A. It is an engineering demonstration, not a claim that one pattern always wins.
In the skills version, the central agent loads a procedure and calls the relevant tools. A separate model no longer interprets every specialist request. Independent agents remain useful where a task needs its own context, reasoning or lifecycle. The example deliberately keeps a research agent alongside the skills.
Our view: architecture should follow the work. Before building a large agent team, identify which components need judgment and which simply execute a well-defined operation. Compare task success, latency and cost using the same cases. Keep authorisation in executable controls; a procedure written in a skill document does not enforce access permissions.
make every extra agent earn its place.
Read the primary source. Editorial interpretation is our own.