Anthropic: a single AI agent with skills often beats a whole team of agents, and is faster and cheaper

Building an AI agent and planning to split the work across several specialist agents? Anthropic, the company behind Claude, now says that a single agent that loads instructions and tools only when it needs them often gives better answers, at lower cost and with shorter waits. Here is when several agents are still the right choice, and how to test.

On 25 October 2026 The Decoder highlighted a new guide from Anthropic: Building agents: When to fold subagents into your main agent loop. In it, Anthropic’s applied AI team writes that it has seen the same pattern across several company projects. A single agent with skills, meaning reusable instructions, often gave better results than both one giant prompt for everything and a group of specialised subagents. It was also often cheaper and faster per task.

The problem with many agents is the handoffs. Every time one agent hands over to another, details get lost, a bit like a game of telephone. When every department builds its own agent, many of them overlap, and the shared instruction becomes a patchwork of conflicting rules. In one example, a company had a dozen specialist agents inside its customer service agent. That made it hard to sound consistent and caused long waits. One customer that compared the two found a single agent was about twice as fast.

Anthropic’s solution is for the agent to see only short descriptions of its skills and tools, and to load the full instruction only when it is needed. According to Anthropic, a clear instruction, 4–5 simple tools for fetching data and firm rules for the answers are often enough. Agents tend to pick tools less well once they have more than 30–50 to choose from. Anthropic itself cut the instruction for Claude Code by over 80 percent for the newer models.

Several agents are still right in some cases, Anthropic writes: when one agent gets measurably worse from too much information, when several tasks need to run at the same time with their own clean context, or when an agent struggles to pick the right tool or has conflicting rules. Anthropic also warns against merging agents blindly. In a test with a support agent the answers got faster, but a process that had worked before broke.

What does it mean for a small business? Start simple. One agent, a clear instruction and a few tools usually go a long way. Skills have also become an open standard that works across several AI services. If you already have several agents, you can use Anthropic’s free migration kit on GitHub. The basic principle is the same whatever tool you use: measure, change one thing at a time and keep what holds up.

Sources

• Anthropic (Claude): Building agents – When to fold subagents into your main agent loop (retrieved 2026-10-26)
• The Decoder: Anthropic says one AI agent beats many subagents now that models can handle complexity alone (25 okt 2026) (retrieved 2026-10-26)
• GitHub: anthropics/single-agent-loop-kit (Anthropics migreringskit) (retrieved 2026-10-26)

TipAI explains AI in plain language for beginners, families and small businesses. Every article has sources and a "How to get started" section.

← All news · Try TipAI