Most AI coding products treat "modes" as personality knobs. You get one chat surface, and a toggle that switches between "Concise" and "Detailed", or a slider for "Creativity". The underlying agent is the same — same tools, same prompts, same output. The slider just nudges the temperature.
Pearl Scout doesn't work that way. The three modes — Ask, Plan, and Agent — are three separate agents. Different tool allowlists. Different system prompts. Different output contracts. The toggle isn't a personality slider; it's a router.
Why we did it this way
The honest answer: because we kept losing trust in the chat surface.
An autonomous agent that can write to your filesystem is great for "implement this feature" and terrible for "explain this error message." If you ask the latter and the agent can write, it will sometimes start writing — opening files, drafting hypotheses, occasionally saving a "fix" you didn't ask for. The slider can't prevent that. The slider is a hint, and the underlying agent is the same agent that, in another context, you want to be aggressive.
The fix isn't to make the agent more cautious. The fix is to make a different agent that cannot do the wrong thing. Ask mode in Pearl Scout has no write tools. It is literally incapable of editing your file. You don't have to trust it not to — the capability isn't there.
What the three modes actually are
Ask is the research agent. It can read files, search content, find by pattern, list directories, and read diagnostics. It cannot write, edit, or run anything. The prompt explicitly tells it: "No TODO lists. No 'let me start by…'. Those phrasings belong to Agent mode. In Ask mode you describe and explain — never narrate work."
Plan is the design agent. It reads the workspace, then emits a structured plan as an interactive checklist. Each phase gets a Run button that re-sends the request in Agent mode, scoped to that phase. The prompt is sharper than Ask: "Issue all the read calls in a single response. Pearl executes them concurrently."
Agent is the ship-it agent. Full tools — read, write, edit, command execution, sub-agent delegation, web fetch. The system prompt is the most opinionated of the three: "Stop only when the task is fully done. Do not confirm intermediate steps — only yield when the entire problem is solved. Three similar lines is better than a premature abstraction."
The trust dividend
Because the modes are real, the in-product UI can be honest about what each one can do. Ask mode says "read-only" and means it. Plan mode says "no writes during planning" and means it. Agent mode shows a one-time confirmation the first time you enter it — because the capability set is genuinely different, and the user should know.
This is what we mean when we say AI should be a first-class part of the editor. The three modes aren't decorations on a single chat; they're three different products sharing a chat surface. The slider model wouldn't survive contact with what developers actually need.
Try it: open Pearl, pick Ask, point it at a file you don't remember writing. Then switch to Agent and ask it to fix the bug you just found. Same chat, two agents.