The Single-Agent Ceiling
A single agent can be remarkably capable. Give it a well-written system prompt, the right context, and a clear task, and it can produce work that would take a human hours. So why would you ever need more than one?
Because a single agent is trying to be everything at once. And everything at once is nothing in particular.
Consider what happens when you give one agent a complex task. Say, evaluating whether your company should launch a new product. That single agent needs to simultaneously research the market, analyze the financials, assess the risks, consider the strategy, and make a recommendation. Five different types of thinking. Each one requires a different cognitive posture. The researcher needs curiosity and breadth. The financial analyst needs precision and skepticism. The risk assessor needs paranoia and imagination. The strategist needs vision and pragmatism. The decision-maker needs conviction and honesty.
What actually happens? The agent does a mediocre version of all five. It researches superficially, analyzes approximately, assesses risks generically, strategizes vaguely, and recommends confidently. The confidence is the problem. The output looks good. It reads like a competent analysis. But it’s shallow everywhere and deep nowhere, because the agent was spread across five incompatible cognitive postures.
This is the single-agent ceiling. It’s not a capability limit. The underlying model is capable enough. It’s an architecture limit. One agent, one context window, one identity, one pass through the problem. The constraint isn’t intelligence. It’s structure.