Anthropic Releases Single Agent Report: “Multiple Sub-Agents Are No Longer Functional”

Anthropic's new report revealed that artificial intelligence models can now solve complex tasks alone, without sub-agents.

As the ability of artificial intelligence systems to solve complex workflows increased, software architects and artificial intelligence researchers were developing different approaches on how these systems should be designed. The dominant view that has prevailed throughout the sector in recent years; The aim was to divide a large and challenging task into multiple small "sub-agents", define specific areas of expertise for each agent and manage them as a network.

However, Anthropic, the artificial intelligence giant behind the Claude model family, announced in its latest comprehensive architectural report that this approach has become obsolete. With the rapid development of large language models (LLM), it has been announced that a single advanced agent is much more performant and reliable than multi-agent networks in directing complex processes.

Paradigm Shift in Artificial Intelligence Agent Architecture

The multi-agent approach, which has long been adopted in the artificial intelligence ecosystem, was born to provide solutions to old generation models with a limited context window and limited reasoning capacity. Developers break down huge coding projects or data analyzes that no single model can solve at once; They were creating separate sub-agents that wrote code, checked the code, searched for data, and fixed errors.

The detailed report published by Anthropic researchers shows that the in-depth reasoning, long context tracking and self-correction capabilities of the new generation big language models eliminate this necessity. Now, an advanced master model can complete the complex task assigned to it on its own by dividing it into logical layers within a single processing flow, without distributing it to more than one subsystem. This situation documents that the concept of “more parts bring better performance” in artificial intelligence software architecture has been replaced by the idea of ​​“a single, more focused, powerful intelligence”.

Weaknesses and Problems Encountered in Multiple Sub-Agent Architectures

The Anthropic report groups the key engineering challenges developers face when building multi-subagent networks, which grow as the system scales, under three main headings:

Communication Errors and Information Loss: While sub-agents exchange data with each other, critical details may be lost during the summarization or transformation of the transferred information. Erroneous input produced by one agent causes all other sub-agents to make incorrect decisions in a cascade (ripple effect).

Uncontrolled Latency and Transaction Cost: Running each child agent independently requires new API calls at each step. This situation not only significantly extends the response time of the system (latency) but also causes a cost crisis by multiplying computing expenses and token consumption.

Error Detection and Control Difficulty (Debugging): When an error occurs in multi-agent systems, it becomes quite complicated to determine which sub-agent, which stage of data transfer or which system command caused the malfunction.

Instead of building such complex layers, Anthropic experts advise developers to; He recommends supporting a single artificial intelligence agent with advanced capabilities with correct system prompts, consistent memory management and well-designed tool integrations. The report emphasizes that the single-agent architecture both minimizes the risk of errors and significantly optimizes the computing budget for institutions.