The more AI agents you add, the worse it gets. Anthropic says so.
Every team builds its own AI agent, someone puts a boss on top, and it gets slower, costlier and worse. It's in Anthropic's own article: one agent with skills "frequently outperformed" a team of subagents on quality, often at lower cost and latency. Handoffs drop details, every team adds rules to the orchestrator until they conflict ("The Frankenprompt"), and your agents end up a copy of your org chart: Conway's law, 1968. Anthropic's advice: one agent, team skills loaded only when needed. In their companion kit, a request across two departments failed twice; with one subagent folded in, it passed twice. But the fold also broke a process, one run out of two. So check which tests flipped, test every chain of agents, and keep a subagent only if it wins.
Source: "Building agents: When to fold subagents into your main agent loop", Claude Resources (Anthropic), Oct 22, 2026 (Diane Sarkis) https://claude.com/resources/articles/building-agents-when-to-fold-subagents-into-your-main-agent-loop Also: anthropics/single-agent-loop-kit (router.md and the recorded run of Oct 8); Melvin E. Conway, "How Do Committees Invent?", Datamation, April 1968.
The + agent button, the meters, the meeting table, the team lanes, the scoreboard, the chain test and the wall of screens are dramatizations; every quote and result on screen is from the sources.
How many agents is your agent?
Transcript
The more AI agents you add, the worse it gets. It's in Anthropic's own article. One agent with skills often beats a whole team of sub-agents on quality, cost, and speed. But, it looks simple. Every team builds an agent, and someone puts a boss on top. So, hand-offs drop details, and every team adds rules to the boss until they conflict. The twist, there's a law for this from 1968. Systems copy their makers' org chart. Your agent did, too. So, Anthropic's advice, one agent with team skills it loads only when needed. Their test, a request across two departments failed twice. Fold one sub-agent in, done. Twice. But, the score hid a break. A simple order check got faster and wrong. One run in two. So, check which tests flipped, not just the score. Test every chain more than once. Keep a sub-agent only if it wins. Fold the rest in, one at a time, and cancel the meeting. Because…
