← All posts LumenAgent

Work mode: what the agent can reach, before it starts

LumenAgent has always had two postures and only ever named one. Work mode is the default one — your documents, your files, your connections — and naming it is what turns “why is it behaving differently?” into something you can look up.

LumenAgent drafting a reply with three generated files attached, beside a timeline of the run

The most common question about any agent is not what can it do — it is what can it see right now. LumenAgent answers that before a run starts rather than in a settings screen three clicks away, and it answers it with a name: work mode.

A mode that exists in the code and not on the screen is indistinguishable, from the user’s chair, from the assistant being moody.

The rule this came from

Two modes, one switch

The studio is in exactly one of two postures, and which one is derived from a single fact: whether a folder on this computer is open as a project. There is no separate setting to disagree with, which is how a mode indicator ends up lying about the tools it actually has.

What each mode is, and what it can reach
ModeWhat it isWhat it can reach
WorkYour documents, your files and your connected appsEvery editor and everything in your library · files and folders on this computer · Gmail, Drive, Slack, Outlook and the rest of your connections · the open web
CodeOne project folder on this computer, with the whole engineering loopSearch and read every file in the project · edit, create and patch with a diff you approve · run the project’s own tests, build and lint · a real terminal, with git as the safety net

The reach is declared, not discovered

Before any work happens, the agent writes a plan naming the sources it intends to use. That is the moment to correct it — the cheapest possible place, because one edit before the run replaces ten edits after it.

LumenAgent — work mode · plan
You

Pull the three supplier contracts out of my Drive, compare the renewal terms, and draft the summary for Priya.

  • Search your library for supplier contracts — 3 files matched
  • Read Google Drive → Contracts/2026 — connection already granted
  • Compare renewal terms · write the summary in LumenDocs
  • Email the summary to Priya — needs your approval to send
4 sources named · nothing sends without you
The plan names its sources before the first step runs. Every line is editable, and the last one is flagged because it leaves your machine.

Where it stops to ask

Reach and permission are two different things, and conflating them is how agents end up either useless or frightening. Work mode describes what is available; the permission gates describe what happens when the agent tries to use it.

  • Reading is free. Your own library, the files you point it at, and the web are read without a prompt.
  • Writing is announced. New files land in your library and are listed in the run log as they are created.
  • Renaming, moving, overwriting and deleting stop and wait — with the exact list of changes shown before any of them happen.
  • Sending and publishing always ask, whatever permission preset you are running, because those are the steps that reach another person.
  • Connections are individually granted. Every service is off until you connect it, and revocable at any time.

Why this is a name and not a checkbox

A checkbox tells you a setting is on. A named mode with a one-line description tells you what the assistant is, and the same sentence appears in the header pill, the context panel and the empty state — so the answer to “what can it reach right now?” is identical everywhere you ask it. That consistency is the feature; the pill is just where you happen to read it.

Questions about modes and reach

How do I know which mode I am in?

The header carries a pill with the mode name. Opening it shows the one-line description and the four things that mode can reach, and offers the single move available from where you are standing.

Can it use a connected app I have not granted?

No. Every service — Gmail, Drive, Calendar, Outlook, Slack and the rest — is off until you connect it individually, and each one can be revoked. Work mode lists connections as reachable; it does not create them.

Does work mode read my whole hard drive?

No. It reads your LumenQube library, and the folders you point it at. Pointing it at a folder is a deliberate act, and the plan names the folder before it reads anything in it.

What happens if a job needs something the plan did not name?

It comes back and asks rather than improvising. The plan you approved is the boundary of the run, which is the reason the plan is written down at all.

Is code mode a separate product?

No. It is the same agent with a different catalog, entered by opening a folder as a project and left by pressing the same pill. See handing off the whole task for the run loop both modes share.

Keep reading