Claude Adviser Strategy: Sonnet Executive + Opus Advisor
Run Sonnet as executive agent handling tools/code/output, consult Opus only as adviser when stuck—beats Sonnet alone on SWE-bench, costs far less than Opus solo, token-efficient for limits.
Adviser Strategy Cuts Token Waste While Boosting Performance
Anthropic's adviser strategy assigns Sonnet as the 'executive' agent for all tool calls, code edits, and user outputs, while Opus acts solely as an adviser invoked only when the executive gets stuck. The adviser reviews context and suggests fixes without writing code or making changes. This outperforms Sonnet alone on SWE-bench benchmarks for both performance and cost, since Opus runs sparingly—only on hard decisions—not every iteration. Sonnet handles routine tasks efficiently with fewer tokens, preserving rate limits (e.g., 5-hour windows fill slower) and context windows versus Opus's high consumption even on simple work. Existing frameworks prioritize app-building over token optimization, but this setup stretches Claude sessions longer within limits.
Trade-offs: Smaller models like Sonnet delegate automatically in Claude Code, but tightened limits demand deliberate efficiency. Invoke via /advisor command with Opus (e.g., 4o as model) alongside Sonnet executive.
Debugging Wins: Sync Fixes and Dependency Resolution
In a real-time sync app built on Sonnet, deletions failed across sessions despite multiple debug rounds. With adviser enabled, Sonnet invoked Opus after prior failures; Opus pinpointed sync logic breaks and exact restructures. Sonnet applied them directly—no extra back-and-forth—fixing deletions even mid-selection across devices.
For major UI library swaps with version conflicts, Sonnet first assessed via Playwright MCP, consulted adviser (which flagged incompatibilities), resolved dependencies sequentially, then redesigned components per advice. Result: Polished, interactive UI, though minor issues lingered. Sonnet skips adviser on minor tweaks (correct behavior) but excels on targeted fixes, saving rounds Sonnet alone would need.
Complex Task Limits: Sequential Execution and Misjudged Scope
Adding a full new page/feature to an existing app saw Sonnet skip adviser entirely, treating it as routine—yielding bugs like cross-component bleed and no auto-sync (required manual 'run' button). Forced nudge invoked Opus, which identified wrong component choices and fixes; Sonnet then enabled streaming edits with proper isolation.
Large refactors took 31 minutes due to Sonnet's sequential handling (no parallel sub-agents like Opus), versus Opus's faster orchestration. For interconnected dependencies or high reasoning needs, adviser doesn't fully bridge gaps—Sonnet picks suboptimal paths, risking more iterations than Opus direct. Model misjudges task complexity, so prompt nudges ensure consultation.
Use for Token-Tight Medium Builds, Skip for Heavy Lifting
Ideal for simpler/medium apps needing occasional deep reasoning amid routine work: Saves babysitting Sonnet, builds more within limits. Avoid for complex apps with failure points—Opus direct saves time despite tokens. Sets realistic ceiling: Helps but requires understanding when to override.