Anthropic Multi-Agent Research System Architecture

Anthropic has implemented a multi-agent research system for Claude that enables the model to perform complex, open-ended research by coordinating a lead orchestrator agent with multiple specialized subagents. This architecture allows Claude to scale its reasoning capacity and token usage, resulting in a 90.2% performance increase on internal research evaluations when using a multi-agent setup (Claude Opus 4 lead with Claude Sonnet 4 subagents) compared to a single-agent Claude Opus 4 system.

Orchestrator-Worker Architecture

The Research system utilizes an orchestrator-worker pattern to move beyond the static retrieval of traditional Retrieval Augmented Generation (RAG). Instead of fetching static chunks of data, the system employs a dynamic, multi-step search process.

The Research Workflow

1. Lead Researcher (Orchestrator): Analyzes the user query, develops a strategic plan, and saves this plan to Memory to ensure persistence if the context window (exceeding 200,000 tokens) is truncated.
2. Subagents (Workers): The lead agent spawns multiple specialized subagents to explore different aspects of the query in parallel. These subagents use search tools, apply interleaved thinking to evaluate results, and return findings to the lead agent.
3. Synthesis: The lead agent synthesizes the subagent findings and determines if further research is required. If so, it refines the strategy or spawns additional subagents.
4. Citation Agent: Once the research loop is complete, a dedicated CitationAgent processes the report to ensure all claims are properly attributed to their sources.

Technical Drivers of Performance

Anthropic's analysis of the BrowseComp evaluation revealed that token usage is the primary driver of performance, explaining 80% of the variance in results. Multi-agent systems effectively scale token usage by distributing work across separate context windows, providing the necessary capacity for parallel reasoning.

Parallelization Gains

To reduce latency, the system employs two levels of parallelization:

• Agent Parallelism: The lead agent spins up 3-5 subagents simultaneously rather than sequentially.
• Tool Parallelism: Subagents execute three or more tool calls in parallel.

These optimizations reduced research time by up to 90% for complex queries.

Prompt Engineering Principles for Agents

Because multi-agent systems introduce coordination complexity, Anthropic utilized specific prompting heuristics to stabilize agent behavior:

• Delegation Guidance: Lead agents are prompted to provide subagents with clear objectives, output formats, tool guidance, and strict task boundaries to prevent duplicated effort.
• Effort Scaling: Prompts include explicit rules to match resource allocation to query complexity (e.g., simple facts require 1 agent with 3-10 tool calls; complex research may require 10+ subagents).
• Search Strategy: Agents are instructed to "start wide, then narrow down," using broad queries initially before drilling into specifics.
• Self-Improvement: Claude 4 models are used as prompt engineers to diagnose failure modes and rewrite tool descriptions, which led to a 40% decrease in task completion time.
• Thinking Processes: The system leverages "Extended Thinking" for the lead agent's planning and "Interleaved Thinking" for subagents to evaluate tool results and identify gaps.

Evaluation and Reliability Challenges

Evaluating multi-agent systems is difficult because agents may take different valid paths to reach the same goal. Anthropic employs a three-tiered evaluation strategy:

1. Small-Sample Testing: Rapid iteration using ~20 real-world queries to spot dramatic impact from prompt tweaks.
2. LLM-as-Judge: A single LLM call evaluates outputs against a rubric covering factual accuracy, citation accuracy, completeness, source quality, and tool efficiency.
3. Human Evaluation: Manual testing to identify edge cases, such as the tendency of agents to prefer SEO-optimized content over authoritative academic PDFs.

Production Engineering

To move from prototype to production, Anthropic implemented several reliability measures:

• State Recovery: Because agents are stateful and errors compound, the system can resume from the point of failure rather than restarting.
• Rainbow Deployments: To avoid disrupting long-running agent processes during updates, traffic is shifted gradually from old to new versions.
• Observability: Full production tracing is used to monitor decision patterns and interaction structures to diagnose why agents fail to find information.

Limitations and Use Cases

Multi-agent systems are not universally applicable. They are poorly suited for tasks with high inter-dependency between agents or those requiring a shared context, such as most coding tasks. They are most effective for high-value tasks involving heavy parallelization and information that exceeds a single context window.

Economically, these systems are resource-intensive; multi-agent systems use approximately 15x more tokens than standard chat interactions and 4x more than single-agent interactions.