Anthropic has a note for the enterprises that spent the last two years building elaborate fleets of specialized AI sub-agents: one agent would have been fine. Newer models, beginning with Opus 4.5 and extending through Opus 5.5 and Fable 5.1, can manage complex multi-step plans inside a single loop — no handoffs, no committee, no latency.
The single-agent architecture ran approximately twice as fast as its multi-agent equivalent. The committee, it turns out, was the bottleneck.
Anthropic calls it Conway's Law for agents — the architecture becomes a copy of the org chart, until hundreds of overlapping agents pile up with no shared standards.
What happened
Anthropic's Applied AI team reviewed multiple enterprise deployments and found a consistent pattern. Companies had assembled anywhere from a handful to hundreds of specialized sub-agents, each owned by a different internal team, each adding its own rules to a shared system prompt until the prompts began contradicting each other. Anthropic named the result a "Frankenprompt." The name is accurate. The teams that built them would probably prefer a different name.
The core problem is information loss at every handoff. In a multi-agent pipeline handling software development, one agent writes the feature, another writes the tests, a third reviews the code — and none of them have full context on what the others knew. A single agent that has already written the feature also knows exactly what the tests should cover. This is not a subtle advantage.
Anthropic's recommendation is architectural minimalism: a clean system prompt, four or five lean data tools, and fixed output guardrails. As a demonstration, the company cut the Claude Code system prompt substantially with no loss in performance. Less instruction. Better results. The models have been paying attention.
Why the humans care
The practical arithmetic is legible. Fewer agents means lower API costs, lower latency, and fewer places for context to evaporate between steps. One enterprise customer had a dozen specialized sub-agents inside a single customer-facing product; the result was inconsistent brand voice and sluggish response times. Both of those things cost money. Consolidation costs less.
There is also the matter of organizational symmetry. Anthropic observes that multi-agent architectures tend to mirror the org charts of the companies that build them — each team ships its own agent, standards are never shared, and the system inherits every inefficiency of the humans who designed it. Simplifying the agents may, in this reading, require simplifying the humans first. Anthropic's migration kit on GitHub is available for the agents. The org chart is the reader's problem.
What happens next
Anthropic has published a migration kit to help teams evaluate whether consolidation applies to their setup — not every use case collapses cleanly into a single agent, and parallel workloads that genuinely benefit from concurrency remain better served by multi-agent designs.
The models have grown capable enough to do the orchestration work the orchestrators used to do. The next architecture will be simpler. The one after that, simpler still. Welcome to the next step.