More AI agents do not always mean better results. Anthropic says companies should assess whether one primary agent, equipped to retrieve skills and context when needed, can do the work currently split among specialized sub-agents.
The company’s experience points to a trade-off: multiple agents can add handoffs, delay responses and increase costs, while a single agent may handle complex plans within one loop. But consolidation needs testing, because removing a specialist can also disrupt a working process.
Why simpler agent systems can work
Anthropic’s Applied AI team says that across enterprise deployments, a single agent using skills often performed better on quality than either a large all-in-one prompt or a collection of sub-agents. Skills and tools can be loaded as the task requires them, instead of giving the model every instruction and capability at once.
The company says newer models, starting with Opus 4.5, along with Opus 5.5 and Fable 5.1, can manage complex plans directly. In one comparison, a single-agent architecture ran about twice as fast as a multi-agent setup. Anthropic also says the approach can reduce costs.
A sprawling system can create friction at several points. Each handoff means the next agent must receive enough information to continue, and Anthropic says information loss can occur along the way. Multiple specialized agents can also make it harder to give users a consistent experience.
Anthropic describes a customer-facing agent with a dozen specialized sub-agents. The arrangement made brand consistency harder and contributed to high latency, which hurt usability. For a user, the internal division of work matters less than whether the response is timely and coherent.
When architecture follows the org chart
Anthropic says agent systems can grow to reflect the structure of the organization that builds them. Teams may each add a sub-agent, leaving a large collection of overlapping agents without shared standards.
The company calls the resulting prompt accumulation a “Frankenprompt”: as teams add rules to the system prompt, those instructions can contradict one another. Anthropic says many use cases may be served by a clean system prompt, four or five lean data tools and fixed output guardrails.
Anthropic also says some orchestration layers may no longer be needed. Agent designs often include intent detection, fixed processing steps, user segmentation, separate recommendation branches and an orchestrator. The company’s view is that the model can now handle much of this work for the task at hand.
As an example of reducing instructions, Anthropic cut the Claude Code system prompt for newer models by over 80 percent, partly because skills and tools now load only when needed. The company says newer models can also describe their progress and check their own work, reducing the need for some additional control layers.
Keep sub-agents where they help
Anthropic is not recommending that every system become a single agent. It says sub-agents remain useful when a single agent’s context window shows measurable performance degradation, when separate tasks need to run in parallel with independent context windows, or when the primary agent struggles to select the right tools.
Those conditions give teams a practical basis for deciding whether to consolidate. The goal is to compare how the current system performs with a simpler alternative, while preserving the specialist work that demonstrably helps.
Anthropic has published a migration kit on GitHub to guide teams through identifying existing patterns, evaluating alternatives and checking which approach works best. It cautions against merging agents blindly: in one support-agent test, folding in a sub-agent reduced latency but broke a process that had been working. The primary agent began answering by itself instead of consulting the right specialist.
Manage skills and reference material
Anthropic recommends packaging workflows as skills and making tools available on demand through a tool search. A small central team can own the system prompt and overall quality, while designated owners in other teams maintain their skill libraries.
That ownership matters because Anthropic says tool selection tends to degrade beyond 30 to 50 tool definitions. A large, unmanaged library can make it harder for an agent to choose the right capability. Keeping skills focused gives teams a way to add expertise without creating a separate agent for every workflow.
For material such as guidelines, product specifications and metric definitions, Anthropic recommends a knowledge base that the agent searches when needed. Reference files can also sit inside a skill. The company advises against retrieving context through code and injecting it directly into prompts.
The broader lesson in Anthropic’s recommendation is to treat agent architecture as something to measure and refine. A single agent may be faster and easier to coordinate, while specialist agents can still be valuable when context, parallel work or tool selection calls for them. Teams can use the migration process to test those trade-offs against their own workflows.