What is agent input? Definition for AI coding workflows
Agent input is whatever an AI coding agent reads to start a task — the user instruction, attached documents, screenshots, and any context the agent fetches on its own.
Agent input is whatever an AI coding agent reads to start a task — the user instruction, attached documents, screenshots, and any context the agent fetches on its own to ground its first move.
It is the starting state for everything the agent does. A vague input ("clean up the dashboard") forces the agent to guess what cleanup means and produces changes that may or may not match the intent. A specific input ("the dashboard chart legend overflows on screens under 1200px — see attached screenshot") gives the agent enough to act precisely.
What makes the input land
The fields a human reviewer would want to see: what is wrong, where is it, what was expected instead, and what does success look like. Format does not have to be elaborate — markdown with the screenshot inline and the source URL on its own line covers most cases.
A visual feedback document built for human consumption is, by default, well-shaped agent input. The discipline that helps a teammate also helps an agent; both fail when the handoff is incomplete.
Frequently asked questions
How is agent input different from agent context?
Input is what the human provides to start the task. Context is the broader pool the agent assembles — input plus retrieved files, plus conversation history, plus any system instructions.
What's the most common reason agent input fails?
Vagueness. A task like "fix the cart bug" forces the agent to guess what's broken; "the cart total shows $0 when a coupon is reapplied — repro on the staging URL below" doesn't.
Capture your first review.
About a minute from open tab to a shareable URL your agent can ingest.
Start capturing