Anthropic reports that single AI agents with on-demand context often outperform multi-agent systems in quality, latency, and cost. The company suggests that newer models, including Opus 4.5, Opus 5.5, and Fable 5.1, are capable of managing complex plans directly within a single loop.
The company observed that multi-agent architectures often suffer from high latency and inconsistent brand experiences. In one internal comparison, a single-agent setup performed twice as fast as a multi-agent equivalent. Anthropic also reduced the Claude Code system prompt by over 80 percent by utilizing on-demand tool loading.
Anthropic has released a migration kit on GitHub to assist teams in evaluating their current agent structures. The company advises against blind consolidation, noting that merging sub-agents can sometimes disrupt established processes. They recommend that a small central team manage the system prompt and overall quality.
The company notes that tool selection performance in models tends to degrade when exceeding 30 to 50 tool definitions. To mitigate this, they suggest maintaining skill libraries and using knowledge bases for reference material. Sub-agents remain useful only for parallel tasks or when context windows face performance limits.
