Anthropic recommends reducing subagents in AI systems
Anthropic recommends reducing subagents in AI systems
Anthropic is advising companies to rethink multi-agent architectures: a single main agent with dynamically loaded skills and tools can be faster and cheaper than complex subagent networks.
Recently, Anthropic has begun recommending that companies reduce the number of subagents, sometimes even merging specialized agents into the main work loop.
Experience from several enterprise projects has shown that an architecture consisting of a single agent and a set of skills is often more effective than complex multi-agent systems. For example, one client increased task execution speed by roughly twofold after moving to a single loop.
The reasons are easy to understand. When tasks are passed from one agent to another, context must be transferred and summarized every time. As a result, some information can be lost, latency increases, and costs rise.
Anthropic now recommends placing planning and task execution in the main agent, which dynamically loads skills and tools as needed. Business processes and domain knowledge should be preserved as reusable components.
However, the article also includes a notable case worth attention. Anthropic specialists once merged subagents in a customer support system. This did reduce latency, but a task that previously worked successfully unexpectedly stopped working. The reason was that the main agent skipped the step of coordinating delivery time with a logistics specialist.
That is why the author emphasizes the need to run end-to-end evaluation (end-to-end Eval) after each subagent merge, paying special attention to which previously successful tasks began failing.
Of course, subagents still remain valuable where parallel research, independent verification, or context isolation are needed.
When designing future multi-agent systems, it is worth asking first: does every subagent really need its own independent context and separate execution loop?
In addition, Anthropic has published a migration toolkit in open source. It allows teams to analyze existing architectures, gradually merge subagents, and compare quality, cost, and latency before and after changes.
Developers working on AI agents should review the toolkit and Anthropic's article.
Why it matters
AnalysisThe recommendation challenges the default assumption that more agents always mean better systems and pushes teams to optimize architecture for latency, cost, and reliability.
Discuss in community
Share your questions and insights with developers
Related in Ecosystem
Practical guides, core terminology, and agent skills
Subagents and Delegation
An architectural pattern for launching ephemeral isolated child agents to execute resource-intensive subtasks in parallel without polluting the parent process's context window.
Anthropic previews Claude Duet, a proactive agent powered by Opus 5.5
Claude Duet lets users delegate tasks to an autonomous agent that can keep working in a virtual machine while they are away.
Claude Code Projects: How a Single AI Orchestrates a Team of Autonomous Agents
A deep dive into Claude Code Projects: Coordinator and Threads architecture, parallel branches, shared memory via MEMORY.md, and cloud agent management.
Claude to Default to Auto Mode for Permissions on Pro, Max, and Team Plans
Anthropic says Claude will begin using Auto Mode by default to manage permissions across Pro, Max, and Team plans.
Subagents and Parallel Execution in Claude Code: Running Multiple Tasks Simultaneously
A comprehensive guide to task parallelization in Claude Code: subagent architecture, shared vs. isolated context, Fan-Out and Swarm patterns, headless CLI orchestration with -p, and token economics.
Claude Haiku (Fast Compact Models by Anthropic)
The most compact and fastest model in the Anthropic lineup, designed for lightning-fast responses, mass text classification, data extraction, and agent routing at minimal cost.