Glossary

Glossary

Essential terms in plain language.

TermDefinition
LLM (Large Language Model)The AI "brain" that reads text and generates text by predicting what comes next based on patterns learned from training data.
Agent LoopThe cycle that makes an agent autonomous: read instructions → decide → act (use a tool) → observe the result → repeat.
ToolA specific capability an agent can use during its loop — browsing a website, reading a file, sending an email, generating an image.
System Prompt (Instructions)The standing orders the agent reads on every turn of the loop — your workflow descriptions, quality standards, and rules.
Context WindowThe agent's working memory — the limited amount of text it can "see" at once, including instructions, conversation history, and tool results.
Step (Primitive)A single, repeatable action in your workflow.
ConstantA part of your workflow that stays the same every time (the template).
VariableA part that changes with each new input (the data).
ParameterizationExtracting the template from your manual process — identifying constants and variables.
ComposabilityConnecting steps so the output of one feeds the input of the next.
Named Workflow (Abstraction)A reusable, packaged sequence of steps with a name you can invoke.
SignalInformation that changes your decision or improves your execution.
NoiseInformation that consumes attention without changing outcomes.
Signal-to-Noise RatioHow much useful information you get per unit of attention spent.
Human-in-the-LoopAI does the mechanical work; you review, decide, and act.
Autonomy LevelHow much independence the agent has for a given step (full, propose-and-approve, or human-only).
FrictionRepeated manual effort that slows you down without adding value.
Knowledge CompoundingThe system getting better over time as workflows accumulate context, patterns, and refinements.
DecompositionBreaking a large workflow into smaller, testable steps.
Exception HandlingRules for what happens when data is missing, ambiguous, or unexpected.
Iterative RefinementImproving results by repeatedly tightening instructions based on real output.
RunbookShort instructions for when and how to run a workflow.