Multi-agent system

A single AI agent that has to do everything becomes unreliable on complex tasks: too many instructions, too many tools, too much chance of a wrong turn. A multi-agent system splits the work. Each agent gets a defined role with its own instructions and tools, and a coordinating agent oversees the whole. How we design such agents and connect them to your systems is covered under having AI agents built.

The advantage is that you can test and improve each role separately (see evals) and that agents can work in parallel. The drawback is complexity: more parts means more places where things can go wrong, and more use of the model. So start with one agent and only split it up when the task calls for it.

For LAB63, five researcher agents work in parallel on brand, market, people, society and media. A strategist agent produces the synthesis and a copywriter agent writes the report. That typically saves three to five days of strategy work per assignment, while the strategist remains responsible for the choices.