Definitions and explanations of key AI terms
An agent is an actor that can sense and interact with its environment. Typically this means the combination of three components: a harness, model, and a set of tools. The harness runs in a loop taking user prompts and forwarding them to the model. The model can ask the harness to use any of the set of tools. These tools are how it interacts with its environment - typically files on your computer.
A Markdown file with a short summary, and longer detailed instructions for a model. Skills can bundle programs along with these instructions, packaging together what to use and how to use it. Skills are offered as tools by the harness to the model, so they are re-usable and shareable ways for models to take actions.
Capabilities of AI tools mentioned in the Chapter 1 recap. Day 3 focuses on five core capabilities of Generative AI, including summarisation and content creation.
The risk that AI perpetuates or amplifies societal biases present in its historical training data, leading to discriminatory outcomes (e.g., in recruitment or loan approval).
The solution to the context window constraint, where large documents are broken into smaller, linkable sections to improve cost efficiency and focus answers.
The limit on the amount of text (context) an AI model can consider at once, functioning as its short-term memory. If a document exceeds this limit, it must be broken into sections (chunking).
One of AI's six core capabilities (introduced in Chapter 1, Day 3) that allows it to interpret questions and generate detailed responses. Deep Research is described as an extension of this capability.
An extension of Conversational Intelligence that actively searches the web in real-time, browses multiple sources, cross-references, and synthesizes findings into structured reports.
A map of meaning (smart tags) that helps AI find the most relevant chunks of data instead of reading everything, thereby saving tokens and cost.
Different business questions require different answer types, including: Yes/No (Classification), A number (Regression), A priority list (Ranking), Natural groupings (Clustering), and A forecast over time (Time-series forecasting).
AI focused on content creation (text, images, code, etc.) using patterns learned from vast data, driven by instructions provided as text (prompts). Its accessibility stems from using Natural Language Processing (NLP).
Garbage In, Garbage Out. States that poor data quality leads directly to poor AI performance, producing confidently wrong outputs that look correct but are flawed.
When AI confidently states plausible-sounding information that is factually incorrect or entirely invented. This is classified as a reputational risk requiring human verification.
A regular program that runs in a loop, accepts user input, sends it to a model, and displays results. A harness might have specific tools available that it can run for the model, and one or more system prompts that alter how the model responds.
An operating model that balances central control (Hub) for governance and standards (Guardrails, LLMOps) with decentralized speed (Spokes) for product delivery and specific use case execution.
LLMOps focuses on the engineering of the model (fine-tuning, version control). AgentOps is the new safety discipline focused on the agent's actions (identity, permissions, audit trails, kill-switches).
Model Context Protocol: a way for agents to gain new tools. A harness is typically an MCP Client, which asks an MCP Server to do things on its behalf.
The gradual degradation of AI performance because the real-world data it encounters (concept drift) or the distribution of inputs (data drift) has changed from its original training data.
The four minimal stages an AI model must follow to remain dependable: Train, Validate, Deploy, Monitor & Improve.
The AI-first mindset where you define what the product learns to achieve (e.g., increase conversion) rather than defining static features.
A structured framework comprising Persona, Task, Context, and Output Specification used to make prompts reliable and precise. This framework ensures the AI understands the audience, constraints, and required format, reducing ambiguity.
The system that combines chunking and embeddings to retrieve the most relevant evidence from a knowledge base before the model generates an answer. This grounds the AI in facts and is used to mitigate hallucination risk.
A systematic scoring framework used to prioritize features based on four components: Reach, Impact, Confidence, and Effort.
A framework used to prioritise AI risks by consequence and timing, categorizing them into: Stop-Work Risks (Priority 1, legal/regulatory threats), Reputational Risks (Priority 2, trust/brand damage), and Operational Risks (Priority 3, quality/cost issues).
Running the same prompt multiple times to compare outputs. Wildly different outputs signal that the prompt lacks sufficient constraints (a prompt quality issue).
Subagents are sub-processes created by an agent which operate with a fresh context and specific configuration. You can define this configuration (allowed tools, custom prompt) in a file.
A progressively rigorous methodology used to verify AI outputs, essential for managing the risk of hallucinations. It consists of: Gate 1: Immediate Scan (checking format/tone); Gate 2: Validation (fact-checking critical claims, testing stability); and Gate 3: Integration (embedding human sign-off and ongoing monitoring).
The small chunks of text (roughly 4 characters each) that AI systems process, representing the unit of cost and the measurement limit for the context window.
Something that a harness can do for a model. This might be 'read file', 'write file', or similar.
Every prompt sent by a harness to a model includes a list of all the tools that are available. If the model decides one of those tools should be called, it sends a specially-formed response to the harness called a 'tool call', detailing what the harness should do on behalf of the model before sending the results back.
AI focused on prediction (e.g., house price, fraud score) based on past trends, often referred to as Machine Learning, and typically focused on numeric datasets.