====== Iterative Deliberation ======
**Iterative Deliberation** refers to a multi-round refinement process in agentic systems where deliberation stages are executed repeatedly, with each subsequent iteration incorporating outputs from previous rounds as context. This technique creates a tunable trade-off between accuracy improvements through iterative refinement and potential degradation from accumulated contextual noise.

===== Overview and Architecture =====
Iterative Deliberation operates as an optional enhancement to staged agentic pipelines, where a deliberation or reasoning stage (typically referred to as "Stage 2" in multi-stage frameworks) can be executed multiple times in succession (([[https://alphasignalai.substack.com/p/how-heavyskill-turns-agentic-harness|AlphaSignal - How HeavySkill Turns Agentic Harness (2026]])).

The process functions as follows: an initial deliberation pass produces outputs that serve as cached inputs for the subsequent iteration, enabling the system to refine its reasoning based on previous conclusions. The framework supports configurable iteration depths, with default settings typically allowing a single pass (N=1) but accommodating up to four rounds (N=4) of refinement, though the optimal iteration count depends on the specific problem domain and computational constraints.

===== Performance Trade-offs =====
The iterative deliberation approach introduces a critical performance trade-off that must be carefully managed during system design (([[https://alphasignalai.substack.com/p/how-heavyskill-turns-agentic-harness|AlphaSignal - How HeavySkill Turns Agentic Harness (2026]])):

**Hit at K (HM@K)**: This metric measures accuracy across all deliberation outputs generated throughout the iteration process. As the system performs additional refinement rounds, HM@K typically rises, indicating that iterative deliberation improves the likelihood of identifying correct answers through repeated examination and reconsideration of the problem space.

**Hit Probability at K (HP@K)**: This metric represents an upper-bound measure of correct answers available across all deliberation outputs. As iterations accumulate, HP@K tends to decay due to the introduction of noise and potential drift in reasoning across successive rounds. Accumulated contextual noise from previous iterations can corrupt or obscure originally valid reasoning paths, reducing the diversity and quality of candidate solutions.

===== Practical Implementation =====
The configurable nature of iterative deliberation allows practitioners to tune the refinement-noise trade-off based on task requirements. Tasks demanding high precision may benefit from multiple iterations (N=3 or N=4), where the improved accuracy from iterative refinement outweighs the noise accumulation effects. Conversely, tasks sensitive to latency or with limited computational budgets may operate with minimal iterations (N=1 or N=2), accepting lower refinement gains to maintain system responsiveness.

The use of caching mechanisms is essential for efficient implementation, as previous deliberation outputs must be efficiently retrieved and presented to subsequent iterations without requiring complete recomputation of earlier stages (([[https://alphasignalai.substack.com/p/how-heavyskill-turns-agentic-harness|AlphaSignal - How HeavySkill Turns Agentic Harness (2026]])).

===== Applications and Limitations =====
Iterative Deliberation appears particularly suited for complex reasoning tasks where multiple perspectives on a problem can improve solution quality. Applications include multi-step planning, logical reasoning, mathematical problem-solving, and decision-making scenarios where reconsideration of premises and assumptions yields benefits.

Key limitations include the computational cost of additional iterations, the diminishing returns in accuracy gains at higher iteration counts, and the risk of reasoning degradation when accumulated noise begins to dominate over refinement benefits. Systems implementing iterative deliberation must monitor performance metrics across iterations to identify optimal stopping points rather than automatically maximizing iteration depth.


===== See Also =====

  * [[iterative_ai_prompting|Iterative AI Prompting]]
  * [[sequential_deliberation|Sequential Deliberation]]
  * [[reflexive_self_correcting_loop|Reflexive Self-Correcting Loop]]
  * [[agentic_workflows|Agentic Workflows]]
  * [[trajectory_selection_strategies_comparison|Trajectory Selection Strategy Performance]]

===== References =====