Your first agent run

The fastest way to understand Arshas is to hand the agent a real task. This guide walks you through your first run: describing a change, watching the plan, and reviewing the diff.

Open a project and press ⌘I (Ctrl+I on Windows/Linux) to open the agent panel. Describe a task the way you would to a colleague: “Add input validation to the signup endpoint and return field-level errors.”

The agent reads your codebase index, locates the relevant files, and proposes a plan — the list of files it intends to touch and what it will do in each. Nothing is modified yet.

Approve the plan and the agent starts working. You can watch each edit stream in, or keep coding — the agent works without stealing focus. When it finishes, you get a single consolidated diff across every touched file.

Review it like a pull request: accept hunks, edit them inline, or reject them. Nothing lands in your working tree until you accept it, and every accepted change is a normal edit you can undo.

Start with bounded, verifiable tasks: add a test, wire up a new endpoint, refactor a function used in three places. Once you trust the loop, scale up to multi-file features and let background agents handle the long-running work.