Wink Pings

Anthropic Multi-Agent Paper: the 90.2% Gain Isn't From the Model

One Opus managing a group of Sonnets beat a single Opus agent by 90.2%. Take it apart: six designs. Look deeper: harness, loop, graph. The problem isn't the model; it's the structure.

Anthropic's multi-agent paper has stunning results.

One Opus was set up to manage a group of Sonnets, and ultimately outperformed a single Opus agent by 90.2%. What got upgraded wasn't the model; it was the structure.

The paper's approach: one Opus handles planning, and a group of Sonnets handles execution. No single conversation window carries the entire workload.

Six designs:

  1. Lead agent: Opus writes the plan to memory before the first search. If the context exceeds 200k tokens it gets truncated, but the plan won't be lost.
  2. Brief: each sub-agent gets the goal, output format, tool list, and 'do not overwrite' boundaries. Without it, two sub-agents will repeatedly search the same area.
  3. Scale effort: a single agent typically calls 3–10 tools; a sub-agent group uses 2–4; deep research uses 10+.
  4. Parallel: run 3–5 sub-agents at the same time, each calling 3+ tools simultaneously. Research time drops by up to 90%.
  5. Self-fixing tools: a test agent repeatedly runs the same broken tool dozens of times, then rewrites the tool's description. Subsequent agents finish 40% faster.
  6. Files + citations: sub-agents write results to files and return paths. Before publishing, another pass checks every citation.

The single agent's problem is: before the question is fully asked, the context runs out first. A team doesn't.

Whitemon packaged these six parts into a prompt; put it into Claude Code and it's usable.

But this isn't a prompt trick. Another Whitemon article proposes a more foundational framework: harness, loop, graph.

Same model, same prompt, different results. Usually the problem is neither the prompt nor the model, but the system outside the model.

  • Harness is the boundary of what the model can do. It includes tools, memory, and permissions.
  • Loop is how the system knows it did the right thing. Don't loop on 'confidence'; loop on 'evidence'.
  • Graph is what to do next. Nodes are independent tasks, and edges are real dependencies.

The paper's six parts are essentially applications of these three layers: Opus writing a plan belongs to harness's memory; brief is a contract; parallelism is graph; self-fixing is loop.

Some readers noted: Anthropic's own documentation says multi-agent consumes about 15x the tokens of normal chat. Others said one manager leading a group of executors—Anthropic has reinvented Indian IT outsourcing.

Ironically, the 90.2% improvement comes not from a stronger model, but from organizational structure.

Your agent keeps failing? Don't rush to change the prompt. First look at which layer is broken.

发布时间: 2026-11-23 05:57