Glossary

Agent carries out tasks using an LLM and tools. In this book it reads and changes code, runs checks, and saves the results of its work.

Task setting explains to the agent what needs to be done and why.

Context includes the instructions, code, history, and other data available to the agent while it works.

Specification describes the system's goal, scenarios, requirements, constraints, and acceptance criteria. The technical approach is described in the plan.

Plan explains how to implement the specification, which parts of the system to change, and how to verify the result.

Context window limits the amount of data the model can take into account in a single call.

Skill saves a recurring procedure as instructions and, when needed, supplements it with scripts, templates, and reference material.

Subagent carries out a bounded part of a task in a separate context on behalf of the main agent.

Oracle gives grounds for judging whether a result is correct. This role can be played by a test, a reference output, or a verifiable user scenario.

Testing seam lets you observe the system's behavior without coupling to internal implementation details.

Tracer-bullet ticket describes a small vertical slice of functionality through the necessary system layers with an independently verifiable result.

Brownfield means working with an existing system and its accumulated constraints. Greenfield means building a new system.

SDD stands for Spec-Driven Development. In this approach an agreed specification guides planning and implementation.

Pattern describes a recurring problem and a way to solve it.

Anti-pattern describes a common mistaken move, its consequences, and a suitable replacement.