What is an AI Coding Agent?
An AI coding agent is a software system that uses large language models to write, modify, debug, and test code. Unlike simple code completion (suggesting the next line), agents operate at a higher level of autonomy — they can take a description of what to build and produce working implementations across multiple files.
Major AI coding agents include Claude Code (Anthropic), Cursor, GitHub Copilot, Windsurf, and various open-source frameworks for building custom agentic workflows.
How agents work
AI coding agents typically operate in a loop:
- Read context. The agent ingests the codebase, documentation, and any specifications provided — via prompt, system instructions, or protocols like MCP (Model Context Protocol).
- Plan. The agent breaks down the task, identifies which files to modify, and determines an implementation approach.
- Execute. The agent writes or modifies code, runs tests, and iterates on errors.
- Verify. The agent checks its output against whatever success criteria it has — if any were provided.
The critical insight: the agent's behavior at every step is shaped by the input it received. Vague input produces vague output. Structured input produces structured output.
The input quality bottleneck
AI coding agents are remarkably capable. They understand language, reason about code, and can implement complex features in minutes. But they cannot:
- Decide what to build. An agent implements what it's told. It doesn't know whether the feature is the right one to build.
- Infer missing constraints. A human engineer asks "what about mobile?" or "what if the API is down?" An agent assumes the happy path unless told otherwise.
- Define success. Without explicit verification criteria, the agent has no way to know when it's done. "It compiles" is not the same as "it solves the user's problem."
This makes the specification — what you hand the agent — the highest-leverage artifact in AI-assisted development. Teams that invest in agent-ready specs get compounding returns. Teams that prompt their way through development hit a ceiling.
Agents and intent engineering
Intent engineering exists because AI coding agents need better input than humans did. A human engineer compensates for a vague ticket with experience, context, and the ability to ask questions. An agent compensates with assumptions — and assumptions compound into bugs.
The IntentSpec format is designed specifically for agent consumption: structured sections that map directly to what the agent needs at each phase of execution. Objectives give the agent context for judgment calls. Constraints prevent unwanted changes. Verification criteria close the loop.
The agent writes the code. The spec ensures it writes the right code.