2026-11-17 · ← News
Anthropic says to dismantle the subagent army. An org chart is not an architecture
Anthropic recommends folding some specialist subagents into a single agent loop with skills and tools. In its deployments that design often improved quality while reducing cost and latency, but the company is reporting customer experience rather than a general benchmark.
The image could not be loaded.
Anthropic argues that many enterprise agents would benefit from fewer subagents and one main loop that loads skills and tools when needed. Ethan Mollick resurfaced the argument as a criticism of corporate harnesses where every team gets its own specialist and another orchestrator sits above them.
One loop can replace specialists that mostly exchange summaries
Anthropic's Applied AI team says a single agent with skills outperformed both a universal prompt and a subagent design across several enterprise deployments. It describes one customer-facing agent with roughly a dozen specialists that struggled to maintain a consistent voice and added noticeable latency. Another customer measured a single loop at roughly twice the speed of its multi-agent harness.
The recommended loop keeps the plan, tool results and subsequent decisions in one context. Procedures load as skills, while live data and actions sit behind tools. Anthropic says one clear system prompt, 4 to 5 narrow data tools and firm output guardrails can often cover most of a use case.
Architecture built around teams turns company boundaries into lossy handoffs
When every department ships its own subagent, the system starts mirroring the org chart. Each specialist has a separate prompt, evals and slice of context. A central orchestrator then receives compressed summaries and has to reconstruct the whole. For tasks that depend on each other or touch the same data, that transfer is expensive and lossy.
Progressive disclosure offers another compromise. The main agent sees short descriptions of skills and tools and loads the full material only when needed. Anthropic also links this approach to a reduction of more than 80% in Claude Code's system prompt. That is a claim about its own product, but it illustrates the direction: fewer rules packed into initial context and more knowledge available on demand.
Subagents still fit parallel work and genuinely separate context
The article does not reject every multi-agent system. Anthropic recommends keeping subagents when one context measurably bloats and degrades performance, when independent tasks can run in parallel or when instructions and tool selection conflict. Separate research branches and black-box verification can still justify decomposition.
The numbers come from deployments and migrations in which Anthropic participated. The company does not publish a shared dataset, models or a complete method for a general architecture comparison. One customer's roughly 2x speedup is therefore not a universal law. It is a reason to measure the local system.
A small end-to-end eval settles more than an argument over diagrams
Anthropic recommends starting with a shared end-to-end eval and says even 10 golden examples can support an initial comparison. A team should fold in one subagent at a time, compare quality, cost and latency and inspect the individual tasks that flipped. In one described test, a regression appeared in only 1 of 2 runs.
The practical signal is not the number of boxes in an agent diagram. It is task completion, repeatability, cost and response time before and after the change. A subagent should earn its place through measurement, not because one team owns a box for it in a slide deck.
Lilith's verdict
When every team sends its own subagent to the meeting, the usual result is another meeting, not a better decision. Every extra chair should be earned by an eval.