Anthropic reports single agents outperform multi-agent setups

Anthropic reports that a single AI agent often outperforms multi-agent setups in quality, cost, and latency. According to The-decoder, the company found that newer models like Opus 4.5 and 5.5 can manage complex plans in a single loop. Anthropic released a migration kit on GitHub to help teams consolidate sub-agents into primary agents that load context on demand. The company notes that in one comparison, the single-agent architecture ran twice as fast as the multi-agent version.

Why it matters

For Swiss businesses, this shift means you can reduce the technical debt of managing complex agent hierarchies. Building a fleet of specialized sub-agents often creates

• higher infrastructure costs
• increased latency for your users
• inconsistent brand voice

You can now focus on refining a single, high-quality system prompt and a lean library of tools. This approach simplifies your maintenance requirements while improving the reliability of your automated workflows.

What to watch

Check your current agent architecture for signs of a "Frankenprompt." If different teams are adding conflicting rules to a single system, your performance will suffer. Audit your tool count now. If you have more than 50 tool definitions, your model may struggle with selection. Move towards a model where a central team owns the core prompt and quality.

Within a year, production environments will shift toward unified agent architectures. This requires developers to prioritise model reasoning over the expansion of complex, nested hierarchies.