learning-systems

Improve swarm task decomposition by running completed task outcomes through an implicit scoring cycle that evaluates execution results without requiring explicit user feedback. This refines future decomposition strategies by translating outcomes directly into actionable learning signals.

Confidence decay is a mechanism that progressively downweights older learning signals to keep swarm pattern evaluation focused on recent task outcomes. This prevents outdated execution results from skewing the active decomposition strategy assessments.

Pattern maturity progression tracks decomposition patterns from candidate to proven stages based on accumulated outcome scores. Ineffective strategies that consistently fail are automatically deprecated, ensuring only proven decomposition patterns are promoted for future planning.

No external dependencies are required beyond the swarm plugin environment itself. The learning system relies on internal components for implicit scoring, confidence decay, and pattern maturity, but does require persistent storage to retain task outcomes, errors, and pattern records.

Trigger the learning cycle immediately after a multi-subtask swarm task completes to record execution outcomes and update pattern scores. Reviewing which decomposition patterns were promoted or deprecated at this point allows you to adjust future planning accordingly.