Glossary

Course concepts. Grows as you work through the modules.

A–Z

AGENTS.md — a file with project instructions, read automatically in every session.

Agent — an LLM in the loop "decision → tool call → result → decision".

Auto (--auto) — a mode that auto-approves actions not forbidden by the rules. Sandbox only.

V

Vibe coding — development where the code is largely written by an agent, while the human sets the task and verifies the result.

Verification — automatic checking of the result (tests, types, linter, run).

H

Hallucination — a plausible but incorrect model output. A consequence of the design, not a "bug".

D

Diff — a line-by-line representation of changes; the main object of review.

Decomposition — splitting a task into independently verifiable steps.

C

Context — everything the model sees in one request: instructions, files, history, the question.

Context window — the maximum amount of context in tokens.

Compaction — automatic compression of the session history when the context fills up.

Cache — a saved stable prefix of the request; reading from cache is 5–10 times cheaper.

L

Lost in the middle — the effect where information in the middle of a long context is used worse than at the beginning and at the end.

M

MCP (Model Context Protocol) — a protocol for connecting external tools to the agent.

Model — a neural network that predicts the next token.

O

Oracle — an automatic source of "yes/no": a test, a compiler, a linter, a run.

P

Permission — an action + resource + effect rule that defines what the agent may do.

Plan-first — the approach "plan first, code second".

Prompt — a text instruction to the model. In this course it is treated as a specification.

R

Reasoning tokens — hidden reasoning tokens in "thinking" models; more expensive, but better on complex tasks.

Review — reading and checking a diff by a human.

S

Session — one working context of the agent; the unit of hygiene.

Spec — a formalized description of a task with verifiable criteria.

Subagent — a separate agent session for a subtask; it does not see the main context.

Skill — a folder with instructions and resources, loaded based on its description.

T

Token — the unit of text for the model (3–4 characters on average).

Tool calling — the mechanism by which the model requests a tool call.

F

Few-shot — a technique with examples in the prompt that set the format or style.

Fast / mini / flash — the class of cheap fast models.

Frontier — the class of the strongest (and most expensive) models.

CI

CI (Continuous Integration) — automatic checks on every commit.

E

Embedding — a vector representation of text; used for search, not for generation.

A

Anchoring — the effect where the model repeats its earlier (possibly wrong) decision.