Enterprise agent diagrams tend to share a silhouette: an orchestrator above, specialists below. Anthropic’s applied AI team argues that shape usually costs more than it earns.
What performed better was a single agent that opens a skill or context only as needed. It beat both sprawling single prompts and subagent networks on quality, and finished one head-to-head at roughly double the speed. Model capability made the difference: Opus 4.5 began holding a complex plan in one loop, and Opus 5.5 and Fable 5.1 push further.
How agent estates rot
The decay is organizational. Team after team ships its own subagent, standards never converge, and hundreds overlap, so the architecture mirrors the org chart. The company dubs this dynamic agent-organized sprawl, its own version of Conway’s Law. Shared prompts then collect contradictory rules: a Frankenprompt.
One customer-facing deployment ran a dozen specialists and paid for it in brand drift and lost context. On-demand skills and tools helped Anthropic trim Claude Code’s system prompt by more than 80 percent in recent releases. Fresh models also self-check their output, retiring some control layers.
When fan-out still wins
Ditching subagents wholesale is a mistake Anthropic warns about: one support test cut latency by folding a specialist in, yet broke a working handoff as the primary started answering instead of routing. Keep them when a context window measurably saturates, when work must run in parallel windows, or when tool selection is shaky.
Otherwise the target is a well-kept prompt and four or five lean tools behind fixed output guardrails, with searchable knowledge bases for reference material. Tool lists blur past 30 to 50 definitions, so owners must keep skill libraries lean; a GitHub migration kit walks teams through comparing options.
