WHY USE MORE THAN ONE AGENT

EXAMPLES Example 1: Customer support chat versus research report. Consider a short customer-support chat in which a shopper asks, "Where is my order?" A single agent can read the question, call an order-lookup tool, and reply in a few steps. Splitting this across three agents would add cost and delay with no benefit. Now consider a complex research report about the history of electric vehicles for a business client. This requires gathering many sources, organizing findings, writing a long document, and verifying every claim. Here a team makes sense: a researcher agent gathers sources, a writer agent drafts, and a fact-checking reviewer agent verifies. Each has a narrower, more manageable job than one agent juggling everything. Example 2: Specialization in action. Suppose the researcher agent's system prompt says, in effect, be thorough and cautious, prefer primary sources, note disagreements between sources, and never state a claim without a source. The writer agent's system prompt says, in effect, write clear and engaging prose for a general reader, use short paragraphs, and use only the findings you were given. If you merged these into one prompt, the model would have to be cautious and bold at the same time. In practice, such a combined prompt tends to produce research that is shallower and writing that is flatter than either specialist would produce alone. Example 3: Context isolation. The researcher agent runs forty searches, and its context grows to tens of thousands of tokens of raw search results, page excerpts, and dead ends. When it finishes, it produces a structured summary of perhaps a page: five key findings, each with a source link and a confidence note. The writer agent receives only that page. Its context contains its instructions and the summary, and nothing else. Because the writer's context is small and focused, its prose is better and its cost is lower. Example 4: Choosing deliberately. A team builds an agent that answers questions about office opening hours. They consider splitting it into a retriever agent and an answering agent, but they run a test and find the single agent answers correctly in under two seconds at a very low cost. Because the single agent is simpler to debug and cheaper to run, they keep it. Months later, the same team builds a tool that writes long compliance reports, and this time they do choose multiple agents, because the single-agent version kept losing track of details in its crowded context. Both decisions were correct, because each was made by looking at the task and the evidence.