ARPInside the ML skill-improvement pipeline.
Our first self-learning example: generate candidates, predict, execute, evaluate, and learn from the evidence.
A skill is a candidate for improvement.
A skill describes how an agent should approach a kind of work. Changing its instructions may improve a result, leave it unchanged, or cause a regression elsewhere. Our first pipeline treats a proposed change as a candidate to evaluate. It retains the current skill as a comparison and keeps a record of the variation and its lineage.
Generate, predict, select.
The loop generates candidate variations. A learned predictor helps estimate their potential and select which candidates to execute. This helps direct the experiment budget toward useful comparisons. A predicted score remains a prediction: it cannot establish that the candidate works or replace an execution result.
Execute, evaluate, ingest evidence.
Selected candidates are executed and evaluated against the stated task. The evidence is ingested for later analysis and training. Valid evaluation requires the intended agent and skill to have actually run. A fallback agent, an unavailable tool, or an empty result can invalidate the environment. Those failures should not quietly become performance scores for the candidate.
Retrain and decide what comes next.
The predictor can be retrained, saved, and reloaded using the accumulated evidence. A comparison between the existing model and a proposed replacement informs the next round. The loop can continue or stop. Versioned candidates, selection evidence, and execution outcomes make the process inspectable. They also help separate a better prediction model from a genuinely better skill.
What this example establishes.
There is an implemented experimental loop to study skill improvement. It provides a concrete first example for Morning’s self-learning direction. It does not yet establish a universal improvement rate, production readiness in every environment, or permission to promote changes automatically. Evaluation and approval remain essential parts of turning an experiment into an improvement a business can trust.
Self-learning systems
Systems that use execution evidence to propose, test, and evaluate improvements. Our ML skill-improvement pipeline is the first example.
Meet Self-learning systems