Adaptive Reasoning

**Judgement:** A well-structured, lightweight heuristic for auto-selecting reasoning depth that is conceptually strong and easy to adopt mentally. It is especially useful in agent systems that explicitly model “reasoning modes” or have expensive long-context/chain-of-thought capabilities. **Strengths:** - Clear scoring rubric with additive/subtractive factors that map naturally onto real query types (multi-step logic, ambiguity, math, high stakes, etc.). - Concrete activation thresholds (≤5, 6–7, ≥8) give straightforward guidance on when to stay fast vs. engage deeper reasoning. - Auto-downgrade rule after complex tasks encourages efficient token usage over a session. **Key Risks / Limitations:** - The prescribed visual indicators (🧠, 🧠🔥) may conflict with environments that disallow emojis or exposing internal reasoning states, and can violate platform-level style or safety policies. - The scoring is heuristic and coarse; some high-stakes or safety-critical queries might be misclassified as low complexity, leading to insufficient deliberation if used naively. - The suggested /reasoning on/off commands and `session_status` tool are not standardized and may not map cleanly onto all hosting runtimes or tool APIs. - No explicit safety layer is defined; it assumes the underlying model/policies will still enforce content and safety constraints even when extended reasoning is activated. **Recommended Scenarios:** - Agent frameworks where developers control or meter “reasoning mode” (e.g., separate fast vs. deep-thinking calls) and want a simple built-in policy to decide which to use. - Mixed workload assistants (support + coding + analysis) where response depth must be tuned dynamically without asking the user. - Systems allowed to reveal reasoning intensity via icons or metadata, to help users understand why some replies are slower or more detailed. **Use with Caution When:** - The hosting platform prohibits emojis or any markers that hint at internal reasoning processes; in such cases, adopt only the scoring logic and thresholds, not the visual indicators. - Handling highly sensitive domains (medical, legal, financial, safety-critical operations) where a dedicated, domain-aware safety and deliberation policy is required beyond this generic rubric.