A tiered framework describing how much an AI system can decide and act without human involvement — spanning from step-by-step human approval to fully autonomous operation where a human can only observe, as a continuous spectrum rather than a binary "autonomous or not."

The mechanisms an agent uses to remember information aren't a single type — they split into several independent categories based on how long something needs to be remembered and what kind of content it is. Understanding this split is what lets you judge what an agent's memory feature can and can't actually do.

Cold Start Problem refers to the situation where an Agent (or a judgment mechanism it depends on) struggles to make precise or robust judgments during an early stage lacking sufficient historical behavior data, since there's no baseline to reference — commonly seen when an Agent has just been deployed, just connected to a new resource, or just started serving a new user.

The upper limit on how much text a model can "see" and take into account at once in a single pass — the <a href="/en/glossary/prompt-techniques/system-prompt/">System Prompt</a>, conversation history, any retrieved documents, and tool descriptions all count toward this same limit together. Once it's exceeded, the earliest content gets pushed out of the window, and the model has no memory that it ever existed.