Claude Advisor: Sonnet Executes, Opus Advises to Cut Tokens
Assign Sonnet as executive agent for routine code tasks and Opus as advisor only for tough spots in Claude Code—saves tokens vs. full Opus runs, outperforms Sonnet alone on SWE-bench, but slower (31min) and buggy on complex UI/feature adds without nudges.
Advisor Strategy Cuts Token Waste by Delegating Roles
Anthropic's advisor strategy in Claude Code flips the default: set Sonnet as the 'executive' agent to handle all tool calls, code edits, and outputs, while Opus acts solely as an 'adviser' invoked only when Sonnet gets stuck. This outperforms Sonnet alone on SWE-bench benchmarks for both performance and cost, and uses far fewer tokens than Opus leading everything—Opus burns tokens even on simple tasks, while Sonnet handles routine work efficiently. Setup: In Claude Code, configure 'advisor' command with Opus (e.g., Claude 3.5 Sonnet executive, Opus advisor). Sonnet delegates to Opus for deep reasoning gaps, avoiding constant high-cost invocations amid tightening rate limits (e.g., 5-hour windows fill faster at peak). Result: Build apps longer within limits, as Sonnet needs less context per task. Trade-off: Existing frameworks prioritize app-building over token optimization, so this custom setup extends sessions without switching tools.
Succeeds on Debugging and Guided Changes, Saves Rounds
In real tests on live apps:
- Debugging sync issues: Sonnet alone failed repeatedly on real-time deletion sync across devices/sessions. With advisor, Sonnet invoked Opus after prior failures; Opus pinpointed sync logic breaks and exact restructures. Sonnet applied fixes in one go—deletions now sync even mid-selection, resolving what Sonnet couldn't alone without extra prompting rounds.
- UI library migration: Sonnet assessed layout via Playwright MCP, consulted Opus on approach. Opus flagged version conflicts first; Sonnet fixed deps, verified app stability, then rebuilt components sequentially per advice. Delivered polished, interactive UI upgrade vs. original.
These show the strategy shines for occasional deep reasoning amid implementation: fewer back-and-forths than Sonnet solo, faster/cheaper than full Opus.
Limitations on Scale: Sequential Execution and Misjudged Complexity
Fails without intervention on bigger tasks:
- Full UI overhaul: Took 31 minutes as Sonnet processed sequentially (no parallel sub-agents like Opus), vs. faster Opus orchestration.
- New feature add: Sonnet skipped advisor entirely, misjudging as 'routine'—introduced bugs like cross-component bleed and manual 'run' needs post-edits. Nudge prompted Opus review: fixed via better component choices and auto-sync. Still needed multiple rounds with real-time state/dependencies.
Sonnet lacks Opus-level foresight for downstream impacts or parallelization, so it picks suboptimal paths. Bottom line: Use for simple/medium apps (straightforward impl + spot reasoning)—build more without babysitting. Skip for complex apps with dependencies/failure points; run Opus directly to avoid time loss from nudges/bugs.