Anthropic tells companies to delete subagents
One agent with skills beat the multi-agent setup, and the advice now is to fold subagents back into the main loop.
Anthropic is recommending that companies remove unnecessary subagents from their AI systems, according to @cherry_mx_reds, who says their team found that a single agent equipped with skills can frequently outperform complex multi-agent systems. One customer saw roughly 2x faster execution after the change.
improving agent architecture is becoming less about what you can add and more about what you can delete
a single agent equipped with skills can frequently outperform complex multi-agent systems
The rules of thumb he lists are a clean division of labour. Skills teach the agent how to do something, covering procedures, expertise and instructions. Tools and MCP give the agent the ability to do things through APIs, actions and live data. Subagents are for parallel work, isolated context, or tasks a single agent struggles with, and not much else.
The mechanism underneath is progressive disclosure. Rather than spawning specialised agents or stuffing everything into one giant prompt, a single agent discovers and loads skills and tools as it needs them. The migration advice is concrete. Fold existing subagents into the main loop, turn their instructions into skills, expose their actions as tools, then run evals to check that performance holds.
Why it cuts against instinct
The usual reaction when an agent system underperforms is to decompose it further, give each piece a narrower job and a dedicated prompt, and hope the orchestration holds. This is the opposite recommendation, and the stated reason is that the models got better at staying coherent over a long single loop, so the scaffolding that used to compensate for that is now overhead.
As @cherry_mx_reds puts it, "improving agent architecture is becoming less about what you can add and more about what you can delete".
Worth being precise about what is evidence here. The 2x number is one customer, relayed secondhand, with no detail on the workload or what was measured. The part you can act on is the eval step, which is the only way to find out whether your own subagent fan-out is earning its latency or just inherited from a time when one agent could not hold the whole task.
Mario Zechner (@badlogicgames) posted a three word response to the news, "no! really?!"
Have the good opinion, not the loud one
One email a day. Under three minutes. What working engineers found out about coding with AI.