Key points
- Anthropic's Applied AI team says one agent loading context on demand often beats multi-agent setups on quality, cost and latency, based on enterprise deployments.
- In one comparison, Anthropic reports the single-agent design ran about twice as fast as the multi-agent version.
- A GitHub migration kit helps teams audit current agent patterns, test alternatives and confirm which design works best.
- Anthropic cautions against merging blindly: in a support-agent test, folding in a sub-agent cut latency but broke a working process.
Why it matters for marketers: Marketing teams running fleets of specialized agents may be paying for extra latency, cost and lost context, and can now test a simpler single-agent design before rebuilding.
Anthropic is recommending that companies check whether their sub-agent architectures can be folded into a single primary agent that pulls in context only when needed. The company says handoffs between many specialized agents add latency, raise costs and lose information. It points to newer models, starting with Opus 4.5, that can run complex plans in one loop.
The advice comes from Anthropic's Applied AI team and its work on multiple enterprise deployments. One customer had a dozen specialized sub-agents inside a customer-facing agent, which made a consistent brand experience hard to maintain and slowed responses. Anthropic describes the pattern as an architecture that mirrors the org chart, with each team shipping its own agent and adding rules to a shared prompt that end up contradicting each other.
For marketing and content teams, that description may sound familiar. Separate agents for segmentation, recommendations or tone can multiply quickly, and each one needs an owner. Anthropic suggests packaging workflows as skills, loading tools through search, and having a small central team own the system prompt and quality. Reference material such as guidelines and product specs should sit in knowledge bases the agent searches when needed.
Consolidation is not automatic. Anthropic says sub-agents still make sense when a single context window degrades, when tasks must run in parallel, or when tool selection becomes unreliable. Its GitHub migration kit guides teams through auditing and testing, and the support-agent example shows why testing matters. Separately, a Vals study cited by The Decoder found that adding more agents raised costs without improving quality.
Sources
This brief is our own summary. Read the full reporting at the source.
