Anthropic says fewer AI agents can deliver better results at lower cost

Some enterprise AI systems may work better after losing their specialist agents. In October 22 guidance, Anthropic says a single agent using skills loaded only when needed frequently beat more complex designs in its deployment comparisons. Its recommendation is to test consolidation—not replace every team of agents with one oversized prompt.

When the agent chart copies the org chart

The pattern starts with a familiar organizational choice: each team builds an agent for its own capability, then a platform team adds a coordinator. These specialist agents, or subagents, have separate instructions, tools and working histories. In the article, Diane Sarkis describes how that arrangement can produce overlapping responsibilities, conflicting rules and lost detail as agents pass work between them.

Anthropic’s Applied AI team says several enterprise deployments benefited from folding planners, routers and domain specialists into the main agent. At one company, roughly a dozen specialists made a consistent brand voice difficult to maintain and added delays that hurt the customer experience.

Keep the history, load the instructions later

The proposed alternative keeps one model working through a task in one context window—the information available to it at a given moment. It plans, calls tools, reads their results and revises its next step. The same model makes the plan and carries it out, rather than passing instructions to specialists and receiving summaries back.

Crucially, the agent does not receive every instruction upfront. It sees short descriptions of skills and tools, then loads the full material when needed. Anthropic calls this progressive disclosure. The company argues that retaining earlier results helps the agent connect findings across tasks, while deferred loading prevents its working history from becoming overloaded.

Simpler does not mean unconstrained. The surrounding software still enforces turn and spending budgets and approval gates. Anthropic also recommends grouping tools by user intent, rather than exposing one tool per API, saying tool selection tends to degrade beyond 30–50 tool definitions in the system prompt.

The guidance preserves three reasons to retain subagents:

• The main agent shows measurable performance loss from too much context.
• Independent tasks need to run in parallel with fresh working histories that can be discarded afterward.
• One agent struggles to choose the right tools or faces conflicting instructions.

A faster answer can still skip a required step

Anthropic recommends building shared end-to-end tests before merging agents; even 10 carefully chosen examples can be a starting point. Teams should consolidate one specialist at a time, converting its instructions into skills and its API calls into tools. Then they should inspect which individual tasks changed from passing to failing, not just the overall score.

Its support-agent example shows why. Folding one subagent into another reduced latency but broke a previously passing process. The main agent used an order-lookup tool and answered without consulting the shipping specialist for the delivery date. The failure appeared in only one of two runs, prompting advice to repeat tests and cover required agent chains and rules.

Consolidation also changes ownership. Individual teams can keep building domain skills and tools rather than separate agents. Anthropic recommends a small central team own the system prompt and end-to-end quality, vetting new skills, tools and tests before release. That replaces isolated measures of each specialist’s performance with shared responsibility for the completed task.

For teams considering that transition, Anthropic provides a migration kit that guides architecture assessment, compares alternatives and verifies the approach. It includes a mock-system demo to try before applying the workflow to a production agent. The decision to consolidate remains tied to testing the actual task.