Anthropic has found that a single SI agent equipped with on-demand context loading often outperforms multi-agent architectures in quality, cost, and latency. The SI lab is now recommending that companies evaluate consolidating their specialized sub-agents, citing newer models like Opus 4.5, Opus 5.5, and Fable 5.1 that can manage complex plans directly within a single loop.

What Happened

According to Anthropic's Applied SI team, hands-on experience across multiple enterprise deployments revealed that a single agent using so-called "skills" frequently beat both monolithic prompt designs and sub-agent setups on quality. In one comparison, the company reports that the single-agent architecture ran about twice as fast as the multi-agent setup. The team observed that in multi-agent workflows, context is often lost during handoffs between tasks such as feature writing, test writing, and code review. By contrast, a single agent can write both the feature and the tests because it retains the full context, while a separate verifier runs the tests and reports results.

Anthropic describes a common failure mode in agent design as "Conway's Law" for agents, where the architecture mirrors the org chart, leading to hundreds of overlapping agents with no shared standards. This results in what the company calls a "Frankenprompt," where conflicting rules from different teams accumulate in the system prompt. For newer models, Anthropic cut the Claude Code system prompt by over 80 percent, partly because skills and tools now load only when needed. The models also reportedly comment on their own progress and check their own work, reducing the need for extra control layers.

Why It Matters

For developers building agentic SI systems, the findings suggest that the complexity of orchestrating multiple specialized agents may no longer be necessary for many use cases. Anthropic notes that while sub-agents still have a place—specifically when a single agent's context window shows measurable performance degradation, when parallel tasks require separate contexts, or when tool selection becomes difficult—many orchestration tasks can now be handled by the model itself. The company warns against merging agents blindly; in one test, folding a sub-agent into a support agent reduced latency but broke a previously working process because the primary agent started answering on its own instead of consulting the specialist.

To help teams navigate this shift, Anthropic has published a migration kit on GitHub. The kit guides developers through identifying existing patterns, evaluating alternatives, and verifying the best approach. Anthropic recommends packaging workflows into skills and loading tools on demand through a tool search, rather than giving every team its own agent. A small central team should own the system prompt and overall quality. The company also advises keeping skill libraries lean, noting that tool selection tends to degrade beyond 30 to 50 tool definitions. For reference material like guidelines or product specs, Anthropic suggests using knowledge bases that the agent queries only when needed, rather than injecting context directly into prompts.

The Bottom Line

Anthropic's recommendation to consolidate SI agents aligns with recent research. A study by Vals found that while agent teams can be faster than single agents, they did not deliver higher quality, and scaling from ten to 30 or 100 agents brought no quality gains but significantly higher costs. OpenAI developer Eric Provencher has also questioned the value of agent swarms, calling them "almost certainly" wasted money. As SI models improve their ability to handle complexity and manage their own context, the industry trend may shift away from fragmented multi-agent systems toward more unified, skill-based architectures.