TL;DR
Anthropic says enterprise teams should test whether newer AI models can handle workflows that previously relied on multiple specialized agents. The company reports that a single agent with skills and on-demand context sometimes improved quality, cost and speed, but warns that consolidation can also break working processes.
Anthropic says companies should evaluate whether newer AI models can do complex work with one primary agent rather than a network of specialized sub-agents, after its Applied AI team found that single-agent designs with context loaded as needed sometimes performed better on quality, cost and latency. The company has published a GitHub migration kit to help teams compare designs and test any consolidation rather than merge agents blindly.
Get business pricing on tech for your team
- Business-only prices and quantity discounts
- Tax-exempt purchasing
- Multiple users, one account, clear invoices
Anthropic’s recommendation draws on its enterprise deployments, according to the company. It says a single agent using skills—reusable workflows or instructions that can be brought in when needed—frequently beat both large, fixed prompts and sub-agent arrangements on quality. In one comparison described by Anthropic, the single-agent architecture ran about twice as fast as the multi-agent setup. The source material does not give the benchmark’s task, sample size or exact cost figures, so the result should not be treated as a general performance guarantee.
The report describes a customer-facing system with a dozen specialized sub-agents. Anthropic said the design made it harder to maintain a consistent brand experience and added latency. In a separate software workflow example, context could be lost as work passed between agents responsible for feature writing, testing and code review. With a single agent, the same system could write the feature and tests using shared context, while a separate verifier ran tests and reported results.
Anthropic says the model can now handle some orchestration that earlier systems assigned to fixed pipelines or separate agents. It recommends a more contained setup: a clean system prompt, a small set of data tools, output guardrails and skills or reference material loaded as needed. The company also says it reduced the Claude Code system prompt for newer models by more than 80%, partly by loading skills and tools on demand. The source describes newer models as better able to plan and check their work, but does not provide independent evaluation data for those claims.
When Fewer Agents Cut Workflow Overhead
For companies building AI assistants, the recommendation challenges the assumption that adding specialized agents automatically makes a system more capable. Each handoff can add latency and coordination work, and Anthropic says it can also cause relevant context to be lost. A simpler design could reduce those costs and make it easier to keep behavior consistent, particularly in customer-facing products.
The advice is not that every agent team should be collapsed into one. Anthropic says sub-agents remain useful when work can run in parallel with separate context, when a single agent’s context window shows measurable performance degradation, or when the agent struggles to choose the right tools. The practical implication is that businesses should compare architectures against their own task and quality requirements, not rely on agent count as a measure of capability.
Products Worth Considering

Anker SOLIX C300 Portable Power Station with 60W Solar Panel, 288Wh, 300W

Logitech MX Keys S Combo Wireless Keyboard and Mouse, Graphite

Bose TV Speaker (2nd Gen) Compact Soundbar
As an affiliate, we earn on qualifying purchases.
How Anthropic Reframes Agent Design
Many agent systems are built around intent detection, fixed processing steps, user segments and separate recommendation branches, with an orchestrator directing work among them. Anthropic argues that some of these layers reflect organizational structure rather than a demonstrated need: teams create separate agents, which can accumulate overlapping instructions. The company describes the resulting prompts as a “Frankenprompt” when additions from different teams conflict.
Anthropic recommends keeping system prompts and quality standards under a small central team, while allowing designated owners to maintain skill libraries. It cautions that tool selection can degrade when definitions grow beyond 30 to 50 tools, though the source does not provide details on how that threshold was measured. For reference information such as product specifications or guidelines, it favors searchable knowledge bases or files attached to skills rather than injecting fetched material directly into prompts.
The report says Anthropic’s conclusions are consistent with a study by Vals, which found that agent teams could be faster than single agents without improving quality; scaling from 10 to 30 or 100 agents reportedly added cost without quality gains. The source provides no study methodology or publication details. OpenAI developer Eric Provencher has also questioned agent swarms, describing them as “almost certainly” wasted money. These are separate observations, not proof that multi-agent designs are generally ineffective.
““Frankenprompt””
— Anthropic, describing a pattern in agent system prompts
Limits of the Reported Comparisons
The source material does not specify the date of Anthropic’s migration kit, the full set of enterprise deployments behind its recommendation, or the methods and task details for the reported speed and quality comparisons. The about-two-times-faster result is one comparison, not evidence that single-agent systems will be faster across workloads. No specific cost savings or numerical quality gains are reported.
Anthropic also recounts a support-agent test in which combining a sub-agent reduced latency but caused the primary agent to answer without consulting the intended specialist, disrupting a process that had worked before. The report does not identify the customer, task, or evaluation criteria. It remains unclear which existing systems can be simplified safely, and how much performance depends on model version, tool design and workflow.
Teams Asked to Test Before Merging
Anthropic’s migration kit on GitHub guides teams through identifying existing agent patterns, comparing alternatives and verifying which approach works best. The company advises against consolidation without testing: teams need to check not only latency, but also whether the agent follows the right workflow, selects appropriate tools and preserves required quality and brand standards.
The source material does not give a schedule for further results or an announced follow-up evaluation. For now, the next step is for companies to run the migration process against their own tasks and compare outcomes before changing production systems. Whether the approach generalizes beyond Anthropic’s deployments will depend on those evaluations and further evidence.
Key Questions
What is Anthropic recommending?
Anthropic recommends that companies test whether one primary agent, using skills and context loaded as needed, can replace some workflows built around multiple specialized sub-agents. It does not advise removing sub-agents in every case.
Did Anthropic prove one agent is always faster?
No. The source reports one comparison in which a single-agent architecture ran about twice as fast. It does not provide enough benchmark detail to establish that result across other tasks or systems.
When does Anthropic say sub-agents are still useful?
Anthropic says they may help when tasks need to run in parallel with separate context windows, when one agent’s context window causes measurable performance degradation, or when the agent struggles to select tools.
What does the migration kit do?
The GitHub kit walks teams through identifying current agent patterns, evaluating alternatives and checking which design works for their workflow. Anthropic warns teams not to merge agents without testing the effects.
Source: rss
