Trajectory Harvester

Software componentModel AdaptationModelsarc:TrajectoryHarvester

A data-flywheel component that extracts successful, and instructive failed, production executions from traces as candidate demonstrations, training trajectories and preference data.

Responsibility. Turns production execution traces into candidate adaptation data.

Also known as: Demonstration curation from production, Data flywheel, Continuous improvement flywheel

Direct neighbourhood (hover for relationship types)

Relationships

invokes dependency

LLM Judge abstract Ch3.5

reads dependency

writes dependency

sends data to dynamic

Annotation Task Router Ch10.3

Design guidance

SHOULD identify successful executions through accuracy metrics and user-satisfaction signals before admitting them as candidates.
SHOULD be monitored for reward hacking and distribution collapse when feeding preference learning loops.
SHOULD collect deployment examples where the model performed poorly or users expressed dissatisfaction and route them for preference labeling.

Quantitative guidance

As stated by the sources; verify before use.

Organizations running continuous improvement flywheels report 25-60% improvements in key metrics versus static approaches (Ch3.5).

Classification

Patterns: Data flywheelIterative demonstration curation
Technologies: NVIDIA NeMo
Quality attributes: Flexibility (ISO/IEC 25010)Maintainability (ISO/IEC 25010)
Risks mitigated: Static agents degrading as distributions shift

Sources

Ch3.5: T. Nguyen, "Prompt Optimization, Few-Shot Learning, Fine-Tuning," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 3.5. ISBN: 9798244538229.
Ch10.3: T. Nguyen, "RLHF Methodology," in Mastering Agentic AI Systems: Guide for the NVIDIA NCP-AAI Exam, 1st ed. 2026, ch. 10.3. ISBN: 9798244538229.