Skip to main content

AI Glossary

Which AI term is your team using as if everyone means the same thing?

Domain dialect. This glossary is the AI-domain subset. The cross-domain canonical lives at /playbook/agents/ (DDL — Dreamineering Domain Language). When this page says "Agent" it means the canonical Digital Agent (DA). When it says "Memory" it must always be qualified with a tier — working, episodic, semantic, or procedural. See Context Flow for the full cascade.

Autonomous Agents​

Autonomous AI Agent​

An autonomous AI software that uses prompts and its surrounding environment to understand, reason, and act toward a defined goal without requiring constant human hand-holding.

Environment​

The sandbox in which an AI agent operates and interacts with external tools. This could be a web browser, a code editor, a database, or an enterprise software ecosystem.

Perception​

The agent's ability to understand and interpret data — text, structured outputs, API responses, or system states — and convert that raw input into actionable understanding.

The Brain​

Large Language Model (LLM)​

The cognitive engine responsible for an agent's thinking and action planning. LLMs serve as the reasoning foundation for most production agents.

Transformer​

The 2017 neural network architecture behind every production LLM. Its self-attention mechanism lets every part of an input relate to every other part simultaneously, removing the context limits of sequential processing. Canonical explanation: How Models Learn.

Next-Token Prediction​

The single training objective of large language models: given the tokens so far, predict the next one. Every broader capability is a byproduct of this narrow objective pursued at scale. See How Models Learn.

Emergent Capability​

An ability a model exhibits that was neither specified in its training objective nor, in documented cases, anticipated by its builders — summarisation, translation, and code generation all emerged this way. Whether a given jump is true emergence or a measurement artifact is contested; see How Models Learn.

Scaling Hypothesis​

The contested proposition that increasing model size, data, and compute will keep yielding new capability, possibly up to general intelligence. A live bet, not a settled fact. See How Models Learn.

Large Reasoning Model (LRM)​

A reasoning-focused language model optimized for complex, context-heavy tasks. Slower than LLMs but delivers significantly higher accuracy — appropriate when correctness outweighs speed.

How Agents Think and Act​

Planning​

The process by which an agent decides the sequence of actions required to reach a goal before executing.

Action​

The actual task executed by an agent in response to a prompt or feedback — clicking a button, writing code, sending an email, querying a database.

State​

A snapshot of the agent's current environment, process, or system condition at any given moment.

Chain of Thought (CoT)​

A reasoning technique where the agent breaks a complex problem into sequential, logical sub-steps before concluding — mimicking how a thoughtful human expert approaches a difficult question.

Reasoning and Acting (ReAct)​

An iterative framework that combines thinking and acting. The agent reasons about what to do, takes an action, observes the result, and adjusts — creating a continuous feedback loop.

Infrastructure​

Tools​

Native or third-party APIs that extend an agent's capabilities beyond its built-in knowledge — enabling actions like web search, database queries, or triggering external workflows.

Memory​

The storage layer that retains both current session context and historical interactions, allowing agents to remain coherent across long tasks and return to prior conversations with continuity. Must always be qualified with a tier: working, episodic, semantic, or procedural.

Knowledge Base​

A curated database from which agents draw domain-specific information to inform and generate accurate outputs.

Architecture​

The structural blueprint of an agentic AI system — defines how all components (reasoning engine, memory, tools, orchestration) interact and function.

Orchestration and Evaluation​

Orchestration​

The end-to-end management of an AI agent's workflow — from receiving an input and reasoning through it to producing and delivering a final output.

Evaluation​

The systematic assessment of an agent's performance and accuracy over time. Without strict evaluation, there is no reliable way to know whether agents are achieving their goals.

Multi-Agent Systems​

When multiple agents are involved, new coordination dynamics emerge.

Multi-Agent System (MAS)​

A framework where multiple AI agents coexist and collaborate within a shared environment, each contributing specialized capabilities to a larger task.

Swarm​

A decentralized form of multi-agent intelligence where agents collectively exhibit intelligent goal-directed behavior through self-organized interactions — no single agent in charge.

Handoffs​

The structured transfer of a task or responsibility from one agent to another, ensuring continuity without information loss.

Agent Debate​

A technique where AI agents engage in structured arguments on a problem to stress-test conclusions and arrive at higher-quality outcomes.

Practice​

Before an AI architecture or workflow review:

  1. Circle every glossary term used in the proposal.
  2. Replace each unqualified use of “agent” with its accountable actor or Digital Agent role.
  3. Qualify every use of “memory” as working, episodic, semantic, or procedural.
  4. Mark any local meaning that differs from this dialect and link its canonical owner.

Proof of done: a new reviewer can explain the actors, memory tiers, tools, orchestration, and evaluation boundary without asking what a key term means.

Context​

  • depends-on Agent Knowledge Contract — owns the cross-domain canonical terms and Digital Agent identity.
  • pairs-with Context Flow — defines how qualified memory tiers move through agent work.
  • applies-to AI Principles — supplies the shared language for the principle routes.
  • applies-to AI Architecture — turns the terms into a system map.
  • proved-by How Models Learn — grounds transformer, next-token prediction, emergence, and scaling terms.

Questions​

Which engineering decision related to this topic has the highest switching cost once made — and how do you make it well with incomplete information?

  • At what scale or complexity level does the right answer to this topic change significantly?
  • How does the introduction of AI-native workflows change the conventional wisdom about this technology?
  • Which anti-pattern in this area is most commonly introduced by developers who know enough to be dangerous but not enough to know what they don't know?

Changes my mind: two informed readers still interpret a glossary term differently after applying its definition to the same architecture.

Next question: which ambiguous term in your current system needs a named canonical owner?