Running the agent

The agent is the core of Arshas: an autonomous coder that takes a task, plans across your repository, edits multiple files, runs your tests, and hands you a reviewable diff.

Open the agent panel with ⌘I and describe the outcome you want. The agent retrieves relevant context from the index, drafts a plan, and shows it to you before touching anything. Plans list every file to be created or modified with a one-line rationale each.

While executing, the agent verifies its own work: it type-checks, runs affected tests, and revises its edits when something fails. You see each step in the activity log, so there is never a mystery about what it did or why.

The agent operates on a scratch layer, not your working tree — nothing is written to your files until you accept the diff. Shell commands beyond your allowlist require explicit approval, and you can stop a run at any point and keep the partial diff.

Treat the agent like a strong teammate: state the outcome, the constraints, and how to verify success. “Add pagination to /api/orders, keep the response shape backward compatible, and update the integration tests” will land in one pass far more often than “fix the orders API”.