Anthropic is urging enterprise AI builders to test whether some specialist subagents can be folded into one main agent equipped with skills and tools. Giving every department its own agent, it warns, can reproduce the company's organizational complexity in software. The org chart doesn't need a digital sequel. Source
The company says that across several enterprise deployments, a single agent with skills frequently outperformed both multi-agent designs and a one-prompt-for-everything approach on quality, often with lower costs and latency per task. Those are Anthropic's reported comparisons, not a guarantee that removing agents will improve any particular system. The article doesn't provide a full benchmark methodology or identify the customers behind those comparisons. Source
Anthropic isn't declaring subagents obsolete. It says it uses them internally and identifies several workloads where keeping them makes sense. Source
Keep the context, lose some handoffs
Subagents are separate agents with their own instructions, tools and context windows—the information available to each model as it works. Skills, by contrast, supply instructions for specialized work, including the procedures and judgment calls that a specialist agent's prompt might previously have contained. Source
Anthropic says keeping connected work in one context lets the main agent connect earlier findings instead of relying on summaries passed back by specialists. It warns that splitting agents along team or problem-type boundaries can produce lost details, overlapping responsibilities, conflicting instructions and evaluations that reward individual subtasks without establishing whether the whole job succeeded. Source
The proposed alternative isn't one enormous prompt. The agent initially sees short descriptions of available skills and tools, then loads their full instructions or definitions when needed. Anthropic calls this progressive disclosure: keep specialized capabilities available without stuffing everything into the model's working context upfront. Source
Planning stays, too. In the suggested design, the same model writes its plan, executes it and revises it as results arrive. The surrounding software still enforces limits such as cost budgets and write-approval gates. Using fewer specialists doesn't mean throwing away the brakes. Source
Comments