Anthropic Advises Shifting From Multi-Agent Fleets to Single-Agent Architectures

Anthropic is advising software teams and enterprise developers to reconsider complex multi-agent frameworks. According to the AI company's Applied AI team, consolidating specialized sub-agents into a single primary agent that retrieves context on demand often delivers superior results in output quality, operational cost, and latency.

The shift is driven by the expanded reasoning capabilities of modern models, including Opus 4.5, Opus 5.5, and Fable 5.1. These systems are capable of executing and tracking complex execution plans inside a single loop, eliminating the need for rigid multi-layered orchestration pipelines. In one benchmark test highlighted by Anthropic, a single-agent setup ran roughly twice as fast as an equivalent multi-agent architecture.

The Cost of Multi-Agent Handoffs

In recent years, many enterprises constructed agent systems mimicking their organizational charts—a dynamic Anthropic describes as "Conway's Law" applied to AI. In these setups, different internal departments build their own sub-agents with dedicated prompts and processing branches.

However, hands-on enterprise deployments revealed significant friction in this approach. In one case, an organization deployed a dozen sub-agents behind a single customer-facing assistant, resulting in high latency and inconsistent brand tone.

Anthropic notes that multi-agent pipelines frequently suffer from context degradation at every handoff between agents, such as transferring instructions from a feature-writing agent to a test-writing agent and then to a review agent. By contrast, a single-agent model can write both the feature and the corresponding unit tests directly because it preserves the full context, requiring only a lightweight verifier to execute tests and log outcomes.

Furthermore, multi-agent sprawl often creates what Anthropic calls "Frankenprompts"—sprawling system prompts loaded with conflicting departmental rules. By streamlining its architecture and loading skills dynamically, Anthropic reduced the system prompt for Claude Code by more than 80 percent on newer models.

Managing Skills and Tool Libraries

Rather than assigning separate agents to individual tasks, Anthropic recommends modularizing workflows into skills that the primary agent loads as needed through a tool-search mechanism.

The company highlighted several best practices for maintaining stability:

• Centralized Prompt Ownership: A small core team should manage the system prompt, output guardrails, and overarching quality standards.
• Tool Count Limits: Model tool-selection accuracy tends to deteriorate once tool libraries exceed 30 to 50 active definitions. Designated team leads should curate skill libraries to prevent bloat.
• Dynamic Retrieval: Background documentation, product specifications, and internal metrics should live in knowledge bases queried on demand rather than being hardcoded or injected directly into system prompts.

When Sub-Agents Remain Necessary

Anthropic cautions teams against merging architectures without careful evaluation. In one customer support test, folding a sub-agent into the primary agent lowered response times but caused the system to answer queries directly rather than routing them to a designated specialist.

To aid developers in evaluating their systems, Anthropic has released a migration kit on GitHub outlining step-by-step assessment workflows. According to the company, multi-agent configurations still provide utility under specific conditions:

• When a single agent encounters measurable performance drops due to an overloaded context window.
• When independent tasks require concurrent execution across isolated context windows.
• When an agent struggles to choose the correct tool from an excessively large set of operations.

Anthropic’s conclusions align with external research on agent performance. A recent study by evaluation platform Vals showed that while agent teams can improve throughput in specific scenarios, scaling from 10 to 30 or 100 agents created sharp cost increases without improving output quality. OpenAI developer Eric Provencher shared similar skepticism toward multi-agent swarms, describing them as "almost certainly" an inefficient use of resources.