About the news
Drawing from enterprise deployments, the company argues that a single primary agent using on-demand skills and tools frequently delivers higher quality, lower cost, and lower latency than layered systems of specialized sub-agents—especially with newer models such as Opus 5.5 and Fable 5.1 that can manage complex plans in one context window.
What the architecture guidance says:
- Prefer one model and one context window running an open-ended tool loop (with turn/cost budgets and write gates enforced by the harness). The model issues batches of tool calls; the harness executes them (often in parallel) and returns results so the same model can decide the next step.
- Use progressive disclosure: keep a clean, short system prompt; package domain knowledge and workflows as Skills (SKILL.md files); load full instructions and tools only when needed via tool search or skill matching. This avoids stuffing the context window with everything upfront.
- Multi-agent handoffs often create latency, higher token costs, and information loss (“telephone game” degradation). In one customer comparison the single-agent design ran roughly twice as fast as the multi-agent harness.
- Sub-agents still have a clear place in three situations: (1) measurable performance degradation from context bloat in a single loop, (2) genuine parallel work that benefits from fresh, discardable context windows, or (3) cases where one agent struggles with tool selection or conflicting instructions.
- Migrate carefully, one sub-agent at a time. Convert specialist prompts into skills and their API calls into tools (or MCP tools), then re-evaluate with the same end-to-end tests. Anthropic released a migration kit that audits each sub-agent and suggests fold / merge / call-as-tool / keep.
- Folding is not risk-free. In at least one support-agent test, latency improved but a previously working process broke because the main agent began answering directly instead of consulting the specialist.
Why it matters
The guidance reflects a broader shift: as frontier models improve at long-horizon planning and tool use, the default design unit is moving from “many specialized agents” to “one capable agent plus a library of skills and tools.” Teams that already run complex multi-agent systems are encouraged to measure first, fold where the numbers improve, and keep sub-agents only where they demonstrably earn their extra context and coordination cost.
