Agentic coding terms, explained

Agentic coding is a way of developing software where an AI agent plans, edits code, runs tools, and checks its own results while the developer sets the goal, supplies context, and verifies the outcome.

Learn more: What is agentic coding?

AGENTS.md is a plain Markdown file in a repository that gives coding agents project instructions, such as how to build, test, and follow conventions. Many agents load it at the start of every session.

Learn more: How to write an AGENTS.md

A token is the unit a language model reads and writes: a short chunk of text, often a word or part of a word. Usage limits and costs for coding agents are measured in tokens.

The context window is the maximum amount of text, measured in tokens, that a model can consider at once. It holds your instructions, the conversation, files the agent read, and tool results.

Learn more: Context engineering for coding agents

Cache reads are tokens the model provider serves from a cache because the same context was already sent earlier in the session. Each agent call re-sends the existing context, which is why cache reads can make up most of the tokens in a long session.

Learn more: Context engineering for coding agents

Compaction is when an agent summarizes a long conversation to free up context window space. It lowers the size of future calls, but the summary can drop details the agent still needed.

Context rot is the informal name for output quality dropping as a session fills with old, irrelevant, or conflicting context. Starting a fresh session for a new task is the usual fix.

A tool call is a step where the agent asks to run a tool, such as reading a file, running a command, or searching code. The tool result is added to the context for the next step.

A subagent is a separate agent session that the main agent starts for a focused task, such as searching the codebase. It works in its own context and returns a short result to the main agent.

Learn more: Subagent orchestration

Orchestration is a setup where a stronger model plans and verifies the work while cheaper or faster models execute the narrow steps. It lowers cost without lowering the quality of the final result.

Learn more: Subagent orchestration

A validation loop is an automated check the agent runs after changing code, such as tests, a linter, or a build, and then uses the result to fix its own mistakes before a human reviews the change.

Learn more: Validation loops for agent-written code

A hook is a script that runs automatically at a defined moment in an agent session, such as before a command or after a file edit. Hooks enforce rules that written instructions can only suggest.