Skip to content

Glossary

Definitions of foundational terms and concepts used across the workshop.

Term Definition
Agent An AI system pairing a foundation model with an operational harness, enabling it to inspect files, execute code, and interact with external environments. See Agent vs. chat.
Agent Harness The client application that manages session state, assembles system prompts, invokes tools, and prompts for human approvals. See Harnesses and models.
Foundation Model The underlying language model that processes input context and generates text or structured tool invocations.
Context The complete sequence of text tokens available to the model on an active turn, including instructions, conversation history, and terminal output. See Context.
Token A discrete sub-word character fragment; the base unit used to measure context window size, throughput, and compute pricing.
Compaction Synthesizing a lengthy session history into a structured summary to free working context while preserving critical analytical state.
Prompt The text instruction, query, or constraint provided to the agent by the user.
AGENTS.md A standardized repository configuration file containing standing instructions and conventions loaded by the agent at session initialization. See Memory.
Skill A packaged operational runbook stored in a dedicated directory, loaded dynamically when a user prompt matches its role.
Tool An executable program, library, or script that an agent can invoke to inspect state or perform deterministic computations. See Tools.
API Application Programming Interface: a structured network interface enabling software systems to exchange data.
Environment The isolated software runtime, package libraries, and system utilities configured for a specific project. See Environment.
Package Manager A tool (such as uv or conda) that automates library installation and pins dependency versions in reproducible lockfiles.
Sandbox An isolated execution boundary (e.g., container or virtual machine) restricting an agent's filesystem and network privileges. See Sandboxing.
Prompt Injection An adversarial attack where untrusted text read by an agent contains instructions intended to hijack model behavior.
Repository (Repo) A project directory tracked by Git, containing the complete history of all tracked files. See Version control.
Commit / Diff An immutable cryptographic snapshot of staged project files / the exact line-by-line differences between revisions.
Secret A sensitive authentication credential, such as an API key, database password, or private certificate. See Secrets.
OCR Optical Character Recognition: converting raster images of text into machine-readable character strings.
BibTeX A standardized text format used for storing bibliographic references and citations.
EDA Exploratory Data Analysis: the empirical practice of profiling distributions, missing values, and anomalies before formal modeling.