The closed learning loop is 📝Hermes Agent's core architectural pattern — solve, document, retrieve, improve, repeat — executed across sessions without human prompting.
Coined by Nous Research, the loop describes the cycle a long-lived agent runs around every non-trivial task. The agent solves the problem, autonomously generates a Markdown skill document capturing the procedure, retrieves that skill the next time the situation recurs, refines or extends it in place when reality contradicts what's written, and repeats — accumulating procedural knowledge in a form portable enough to share across runtimes via the 📝agentskills.io standard.
The loop is the architectural counterpoint to two adjacent patterns. Unlike training-time reinforcement learning, where the model's weights change, the closed learning loop changes the agent's skill library — its operating knowledge — without retraining the underlying model. Unlike retrieval-augmented generation, where the agent reads static documents a user has provided, the loop produces its own documents from lived experience and edits them when they go stale. The agent gets more capable the longer it runs because procedure improves, not because parameters change.
The threshold for capture is heuristic rather than fixed: Hermes typically generates a skill after five or more tool calls, the rough boundary for "complex enough to be worth documenting." Skills are stored locally, version-controlled, exportable, and contributable upstream to the agentskills.io community marketplace. The architectural claim is that compounding skill libraries produce a different kind of intelligence than either bigger models or longer context windows — one that emerges from operating in the world rather than from inference-time compute.