Anthropic Publishes Guidance on Consolidating Multiple AI Agents Into One, With Execution Time Roughly Halved in One Case

On October 22 local time, Anthropic published design guidance and a migration kit for consolidating some of the subagents used in enterprise AI systems into a single agent. The target audience is companies that have split their work across multiple AI agents, meaning AI systems that carry out tasks on their own. Anthropic suggests first assessing whether the work can be consolidated into one agent that pulls in the information it needs as it goes.

What changes when you consolidate

The new guidance centers on a single main agent that owns a workflow from start to finish. Within one flow, it makes a plan, uses tools, then looks at the results and decides what to do next. Work procedures are packaged into “skills” that get loaded when needed, while data lookups and execution are handed off to tools. The number of back-and-forth exchanges the model makes (turns), cost ceilings, and the conditions that require human approval are managed by the program that runs the AI.

Anthropic said that across deployments at a number of companies, a single agent using skills frequently came out ahead on response quality. Cost per task and waiting time also went down in many cases, the company said. In one customer’s architecture comparison, execution time fell by roughly half. Anthropic added that it also cut the default system instructions in Claude Code for advanced models by more than 80%.

When you should still run multiple agents

Anthropic has previously published performance results for approaches in which multiple agents work together. The company said that in 2025, its own research system scored 90.2% higher than a single agent on an internal evaluation that involved investigating several directions at once. This latest guidance recommends measuring the architectures for yourself, task by task, and deciding on that basis. Work that requires running unrelated investigations in parallel qualifies as a case for leaving subagents in place.

The other criteria that determine whether to consolidate are the volume of information, tool selection, and conflicting instructions. Anthropic recommends keeping subagents in three situations: when the volume of information handled at once becomes large enough that performance actually degrades, when the agent has trouble picking the right tool, and when instructions inside the system prompt clash with each other. The company stressed that the same evaluation tasks should be used before and after consolidation. Among the published cases was one in which waiting time dropped but existing work procedures broke down.

How Korean companies can use this

The migration kit released on GitHub is a set of implementation examples for companies evaluating their existing agent architecture. It includes sample checks that can be run without connecting a live model and without incurring any cost, while attaching a model for evaluation may incur usage charges. The repository states that it is not maintained and does not accept outside contributions. Companies considering adoption should look at how they will manage it afterward as well as at evaluation costs.

Anthropic has officially stated that Claude use and developer activity are taking place in South Korea as well. The Claude web and desktop apps support Korean as a display language. This guidance and kit are focused on reviewing the design of enterprise AI that is already in use. Korean companies can use them as a reference for picking which of the tasks they have divided among multiple AI agents to test for consolidation.

Anthropic’s proposal amounts to deciding how many AI agents to assign work to based on actual results. What adopting companies need to watch is not just speed gains but response quality, cost per task, and whether existing procedures still run as before. Comparing before and after with the same tasks shows whether the work is handled better once it is consolidated into a single agent.