Claude Developers Share Two Multi-Agent Patterns: Advisor and Orchestrator

Claude developers posted details on two multi-agent setups, Advisor and Orchestrator, that combine Sonnet 5 and Fable 5 to raise task accuracy without matching the cost of running the larger model alone.

The post from X user @shao__meng summarized internal usage at Anthropic. Advisor routes guidance requests from Sonnet 5 to Fable 5 through tool calls. Orchestrator lets Fable 5 break work into subtasks and assign them to multiple Sonnet 5 instances. Both patterns were tested on SWE-bench Pro and BrowseComp.

Advisor pattern lifts Sonnet 5 accuracy at moderate added cost

In Advisor mode, Sonnet 5 handles execution while making targeted tool calls to Fable 5 for planning or verification steps. On SWE-bench Pro with 482 tasks, this raised pass rate from 75.5 percent at 0.75 dollars to 84 percent at 1.40 dollars. Fable 5 used alone reached 91.5 percent at 2.25 dollars. The hybrid approach captured roughly 92 percent of that performance at 63 percent of the cost.

The pattern keeps most inference on the lighter model and uses the heavier model only when the executor signals uncertainty or needs high-level direction. The post indicates this routing occurs dynamically through function calls rather than fixed hand-offs.

Orchestrator pattern improves BrowseComp results with parallel workers

In Orchestrator mode, Fable 5 acts as planner and dispatcher. It decomposes questions, assigns subtasks, and merges results from several Sonnet 5 workers. On BrowseComp, the setup scored 86.8 percent at 18.53 dollars. All-Sonnet 5 reached 77.8 percent at 16.01 dollars. All-Fable 5 reached 90.8 percent at 40.56 dollars. The orchestration delivered about 96 percent of full-Fable 5 accuracy at 46 percent of its cost.

This approach trades some added coordination overhead for reduced reliance on repeated calls to the largest model. The post notes that task division quality directly affects final accuracy, and Fable 5's planning step is the critical variable.

Cost-performance tradeoffs favor hybrid routing over single-model runs

Both patterns target the same practical problem: Fable 5 delivers higher standalone accuracy but at more than double the inference cost of Sonnet 5. Pure Sonnet 5 runs leave accuracy on the table. The hybrids close most of the gap without paying full price.

The Advisor pattern shows the largest relative lift on coding-heavy SWE-bench Pro tasks. The Orchestrator pattern shows clearer gains on BrowseComp, where information retrieval and synthesis benefit from parallel worker execution. Neither pattern matches the top line of the largest model alone, yet the cost savings reach 37 to 54 percent depending on benchmark.

Questions remain about generalization and failure modes

The reported numbers come from internal Anthropic testing on two specific benchmarks. No independent replication data has appeared yet. It is unclear how the patterns behave on tasks outside SWE-bench Pro and BrowseComp or when worker models receive conflicting instructions.

The post does not detail error recovery when Fable 5 planning contains mistakes or when Sonnet 5 workers return inconsistent outputs. These edge cases could narrow the observed gains once the setups move beyond curated test sets.

Next signals to watch

Teams that adopt either pattern will track three concrete indicators over the next three months. First, whether public benchmarks adopted by other labs begin to publish hybrid-model scores alongside single-model baselines. Second, whether error rates on production coding or research tasks drop measurably after introduction of dynamic routing or orchestration. Third, whether Anthropic or competing labs release open tooling that implements the same Advisor or Orchestrator call patterns.

Each indicator will show whether the reported accuracy-cost balance holds outside controlled evaluations.