Anthropic says single AI agents beat subagent fleets in speed and cost

Anthropic attributes the rise of complex multi-agent systems to "Conway's Law" for agents, where architecture mirrors the org chart, leading to overlapping agents and contradictory "Frankenprompts." The company suggests that a clean system prompt, lean data tools, and fixed output guardrails are often sufficient. For instance, Anthropic cut the Claude Code system prompt by over 80 percent for newer models because skills and tools now load only when needed.

While sub-agents remain useful for parallel tasks or when context windows degrade performance, Anthropic warns against blind consolidation. In one test, merging a sub-agent broke a previously working support process. To guide teams, Anthropic published a migration kit on GitHub. These findings align with a study by Vals, which found that scaling agent teams from ten to 100 increased costs without improving quality, and with OpenAI developer Eric Provencher, who called agent swarms "almost certainly" wasted money.