Anthropic’s multi-agent research system raises the bar for open-ended AI reasoning
Jun 18, 2025
On June 13, Anthropic published a detailed account of how it built its new multi-agent research system—the foundation for Claude’s recently launched Research feature. The post outlined both the technical architecture and the design philosophy behind using multiple AI agents to perform complex, open-ended tasks.
The system reflects a broader shift underway in AI development, as leading organizations move from single-model architectures toward orchestrated, multi-agent systems designed to handle more open-ended and research-intensive tasks.
This marks a move away from monolithic models and toward collaborative agents
Anthropic’s system moves beyond the traditional single-agent approach. Instead of asking one model to tackle a broad and nuanced query alone, Claude assigns different parts of the task to multiple specialized subagents.
These subagents then work in parallel—querying sources, interpreting data, and synthesizing insights—before feeding their results back to a lead agent that crafts the final response.
Anthropic found that this multi-agent architecture dramatically outperformed single-agent methods for tasks that required exploration across a wide set of sources. For example, identifying board members from S&P 500 companies in the IT sector is a search space too large for one model to handle effectively.
By parallelizing research, the system achieved up to 90% better performance on breadth-first tasks.
Prompting, not just modeling, defines the system
Key to Anthropic’s system is prompt engineering—not as a set of manual hacks, but as a structured design layer. The team relied on prompts to define the role of each agent, guide its behavior, and ensure consistency across tasks.
Even subtle changes in phrasing had cascading effects, which made prompt clarity and testing as critical as model quality.
The company also emphasized the importance of tool reliability. Subagents interact with tools like search and citation engines, so clear documentation and predictable outputs were vital. Anthropic even deployed LLMs to evaluate and refine tool descriptions and interfaces, reducing latency and improving performance.
Production-grade orchestration comes into focus
Anthropic’s post also highlights the engineering complexity of deploying a multi-agent system in a production environment. The system includes checkpointing, retry logic, and rainbow deployments for safety.
It currently operates synchronously, but asynchronous orchestration—where agents operate and respond at different times—is under development.
To evaluate performance, Anthropic uses both automated and human review methods. LLMs assess factors like accuracy and source quality, while human evaluators catch subtle issues, such as overreliance on SEO-optimized content.
Observability is essential: engineers monitor not just outcomes, but how prompts and agent behaviors evolve over time.
Anthropic’s approach highlights key implications for the broader AI ecosystem
Anthropic’s work has several important takeaways for AI developers and enterprise leaders:
The era of single-agent AI is giving way to collaborative systems that mirror human workflows.
Prompt engineering and orchestration strategies are becoming core parts of system design, not just peripheral concerns.
Tool clarity and protocol design are emerging as new quality benchmarks for multi-agent architectures.
Observability and human oversight are essential safeguards as systems grow more complex.
Token and compute costs remain high. These architectures only make sense for high-value use cases with complex reasoning needs.
Together, these shifts signal that building effective AI is about designing intelligent systems that can coordinate, adapt, and operate with accountability at scale.
Centific is already exploring multi-agent AI systems
At Centific, we’ve been exploring many of these same questions. In our recent article, “Model-Context-Protocol can improve AI adoption—if you take the right steps,” we outline how coordination protocols and memory management are essential to building trustworthy multi-agent systems.
As the industry continues its shift toward multi-agent AI, Centific remains focused on helping enterprises build frameworks that are technically sound, scalable, and grounded in human judgment. Centific’s frontier AI data foundry platform plays a critical role in this effort by providing the high-quality, context-rich data and human-in-the-loop workflows needed to train and evaluate agent-based systems with precision.
Learn more about Centific’s frontier AI data foundry platform.
Share
Categories
Anthropic
Multi-agent systems
Agentic AI
Reasoning AI
Latest news
Nvidia’s momentum is fueling a stronger AI ecosystem
Nvidia is transforming the AI landscape. Learn how its ecosystem is accelerating innovation across industries and domains.
Jun 3, 2025
OpenAI’s move into hardware could change how AI is trained
The AI interface is shifting from screens to space. Explore how OpenAI’s hardware ambitions are redefining how AI is trained, evaluated, and experienced.
May 28, 2025