Anthropic Recommends Consolidating Sub-Agents into One: Architecture and Benchmark Review
Anthropic recommends replacing multi-agent systems with a single agent using lazy skill loading — faster, cheaper, and higher quality.
Anthropic researchers found that one primary model with on-demand context loading often works faster, cheaper, and better than a network of specialized sub-agents. As The Decoder reports, the company is preparing a migration toolkit on GitHub so teams can test this approach on their tasks. For business owners implementing AI in support, sales, or internal processes, this is a signal to rethink architecture: complex multi-agent schemes may add unnecessary costs and latency without quality gains.
Previously, the standard scheme was: intent detector, fixed pipeline, user segmentation, separate recommendation branches, and an orchestrator on top. New models — starting with Opus 4.5, Opus 5.5, and Fable 5.1 — can build and execute complex plans in a single cycle. According to The Decoder, in one comparison the unified architecture performed roughly twice as fast as the multi-agent one. Companies that have already deployed dozens of sub-agents face growing latency, context loss during handoffs, and dilution of brand voice.
What changed: models learned to hold complexity in one cycle
Anthropic notes that modern models can manage multi-step processes without an external orchestrator. They themselves decide which tool to call, when to load a reference, how to verify their result, and when to stop. The publication does not specify the exact capabilities of Opus 4.5, Opus 5.5, and Fable 5.1, but the emphasis is that “skills” and tools are loaded on demand, not all at once.
This changes the economics: instead of paying for tokens to support dozens of agents and pass context between them, you pay for one cycle with dynamic loading. Per Anthropic's experience, the Claude Code system prompt for new models was reduced by more than 80 percent — precisely due to lazy loading of skills and tools. Models also comment on their reasoning and self-verify, making many control layers that teams added on top redundant.
Why multi-agent schemes degrade quality and increase maintenance costs
According to The Decoder, Anthropic's Applied AI team observed the same problem across several enterprise deployments: agent architecture mirrors org structure. Each team ships its own sub-agent, resulting in hundreds of overlapping entities without unified standards. The result — a “frankenprompt”: the system prompt accumulates rules from different teams, they contradict each other, quality drops, and debugging becomes impossible.
Every handoff between sub-agents means information loss, added latency, and error risk. In one example, a company maintained a dozen specialized sub-agents inside a customer bot: keeping a unified brand experience was hard, and response delays killed usability. Anthropic calls this “Conway's law for agents” — the system structure repeats the organization's structure, not the task.
When sub-agents are still needed
Anthropic does not propose abandoning them entirely. As The Decoder writes, sub-agents are justified in three cases: if a single agent's context window causes measurable quality degradation; if multiple tasks must run in parallel in separate context windows; if a single agent gets confused choosing tools.
The risk of blind merging is also real: in a test with a support agent, embedding a sub-agent reduced latency but broke the workflow — the main agent started answering itself instead of calling the right specialist. Migration requires verification at each step, not just merging code.
How to migrate: Anthropic's toolkit and approach
The company published a migration toolkit on GitHub that guides teams through a workflow: identify current patterns, evaluate alternatives, verify the best approach. The key recommendation — package workflows into skills and load tools on demand via tool search.
A small central team should own the system prompt and overall quality. Separately, Anthropic warns: model tool selection quality drops after 30–50 definitions, so each team needs a skill library owner to prevent bloat. Reference materials — guidelines, product specs, metric definitions — are better placed in knowledge bases the agent queries on demand via a search tool. Injecting context via code directly into the prompt is not recommended.
What this means for small business in Russia and CIS: concrete steps
Many small companies in the region are already trying AI agents in tech support, order intake, lead qualification, or internal help. If you have one bot working — check whether you've layered prompts from different employees onto it (“sales manager wants this,” “legal requires that”). This is a direct path to a “frankenprompt.”
If you already have multiple bots or sub-agents (one for FAQ, another for order status, third for upsell), evaluate: do you need an orchestrator and context passing? Often simpler: one agent with a skill set: “check inventory,” “search knowledge base,” “process return.” The skill loads only when the model decides it's needed.
Plan for this week:
- Take inventory: how many agents, sub-agents, prompts, who changed them in the last month.
- Calculate average response time and error rate on agent handoffs (if logs exist).
- Try building one agent with lazy loading of 3–5 skills on a test dialogue set. Compare latency, token cost, and answer quality against a checklist.
- Assign one owner for the system prompt and skill library — even if the team is two people.
- Move reference materials (pricing, return policy, regulations) into a search index accessible to the agent via a tool, not into a hardcoded prompt.
Anthropic's GitHub toolkit is open — you can use their evaluation checklists without writing code from scratch. The publication does not specify whether the toolkit supports Russian-language models or only the Anthropic ecosystem, but the logic is portable.
Conclusion
Studies referenced by The Decoder — including work by Vals and the opinion of OpenAI developer Erik Provencher — converge: scaling agents beyond a dozen yields no quality improvement but significantly increases costs. For business this means: before hiring developers to maintain a herd of agents, check whether one strong model with dynamic context loading handles it. Most often — yes.
Frequently Asked Questions
Why does Anthropic recommend combining sub-agents?
Anthropic discovered that combining sub-agents into a single agent with lazy skill loading leads to faster speed, lower cost, and improved task quality. This is due to optimized interaction and reduced redundant calls.
What is “lazy skill loading”?
Lazy skill loading is an approach in which required functions or knowledge are activated only when they are actually used. This allows the agent to stay lightweight and efficient by loading only the components needed for the current task.
What advantages does a single agent offer over a multi-agent system?
A single agent delivers higher task execution speed by eliminating the need to coordinate between sub-agents. It also reduces operational costs through optimized token usage and API calls, and provides more stable result quality.
- Meta Platforms Inc. activities (social networks Facebook and Instagram*) are recognized as extremist and prohibited in the Russian Federation.
Web development, digital advertising and marketing expert since 2012. Founder of AIFORSALE.