OpenAI lands in roughly the identical place. Its sensible information recommends maximizing a single agent’s capabilities first as a result of one agent plus instruments retains complexity, analysis, and upkeep extra manageable. It explicitly suggests immediate templates as a option to take up branching complexity with out leaping to a multi-agent framework. Microsoft is equally blunt: If the use case doesn’t clearly cross safety or compliance boundaries, contain a number of groups, or in any other case require architectural separation, begin with a single-agent prototype. It even cautions that “planner,” “reviewer,” and “executor” roles don’t robotically justify a number of brokers, as a result of one agent can usually emulate these roles by way of persona switching, conditional prompting, and power permissioning. Google, for its half, provides a very helpful nuance right here, warning that the fallacious alternative between a sub-agent and an agent packaged as a instrument can create huge overhead. In different phrases, typically you don’t want one other teammate. You want a operate with a clear contract.
Microsoft makes yet one more level that deserves further consideration: Many obvious scale issues stem from retrieval design, not structure. So, earlier than you add extra brokers, repair chunking, indexing, reranking, immediate construction, and context choice. That isn’t much less formidable. It is extra grownup. We discovered this the laborious manner with microservices. Complexity doesn’t vanish if you decompose a system. It relocates. Again then, it moved into the community. Now it threatens to maneuver into hand-offs, prompts, arbitration, and agent state.
Distributed intelligence remains to be distributed
What may have been one robust mannequin name, retrieval, and some fastidiously designed instruments can shortly flip into agent routing, context hand-offs, arbitration, permissioning, and observability throughout a swarm of probabilistic parts. Which may be value it when the issue is actually distributed, however usually it’s not. Distributed intelligence remains to be distributed programs, and distributed programs aren’t low cost to construct or preserve.