# Anthropic suggests consolidating sub-agents into one agent with on-demand context

published: 2026-10-25T11:05:29.788Z
topic: frameworks
source: https://the-decoder.com/anthropic-says-one-ai-agent-beats-many-subagents-now-that-models-can-handle-complexity-alone/
url: https://aiagentnews.io/stories/anthropic-suggests-consolidating-sub-agents-into-one-agent-with-on-demand

Anthropic says a single agent that loads context as needed often beats a group of specialized sub-agents on quality, cost and latency. It has published a migration kit on GitHub to help teams evaluate consolidating their agents.

Anthropic's Applied AI team reports this from hands-on work across several enterprise deployments, where a single agent using skills frequently outperformed both monolithic prompts and sub-agent setups on quality. In one comparison, the single-agent design ran about twice as fast as the multi-agent setup. One company had a dozen specialized sub-agents inside its customer-facing agent, which made brand consistency hard and added high latency.

The company argues that multi-agent pipelines lose context at each handoff, and that newer models can handle orchestration themselves. Anthropic cut the Claude Code system prompt for newer models by more than 80 percent, partly because skills and tools now load only when needed. It says sub-agents still make sense when a context window shows measurable degradation, when tasks must run in parallel, or when tool selection struggles.

Anthropic warns against merging blindly: in one support-agent test, folding in a sub-agent cut latency but broke a working process, as the primary agent began answering instead of consulting the right specialist. It recommends packaging workflows into skills, loading tools through tool search, keeping a small central team in charge of the system prompt, and watching skill libraries because tool selection reportedly degrades beyond 30 to 50 tool definitions. A study by Vals reached similar findings, with more agents adding cost but not quality.

## Why it matters if you build agents

Developers running several specialized agents may find that fewer, leaner agents with on-demand skills and tools are cheaper and faster to operate. Before merging agents, they should test for context lost at handoffs and for regressions such as a primary agent skipping a specialist. Keep tool counts in check, since selection quality reportedly drops past 30 to 50 definitions.