The gap between an assistant and an agent is small to describe and enormous in practice. An assistant tells you how to build the board pack. An agent builds it, in your brand, using your numbers, and hands you the files. Lumen Agent is the second kind — and the whole design question becomes how does it stay trustworthy while doing that.
Let it work. Keep the last word.
The rule the whole product is built around
The loop
Every job runs the same five stages. The third one is a gate, not a step: reading, searching and analysing run freely, but anything that writes a file, sends a message or spends credits waits for you.
A real job, traced end to end
Here is a job that would normally cost most of an afternoon and touch four different applications: quarter-end numbers in a spreadsheet, last quarter’s deck, a written commentary, and an email to the board. One sentence starts it.
Take Q3-actuals.xlsx, build the board pack in our template like last quarter, write the commentary on anything that moved more than 10%, and draft the email to the board with it attached.
I can do all four. Before I touch anything, here is the plan.
I will use Q3-actuals.xlsx from your library and Board pack — Q2.lslides as the template reference.
- Read Q3-actuals.xlsx · reconcile against Q2 · flag every line that moved >10%
- Build 14 slides from the Q2 template, in the brand kit, with the new figures
- Write the commentary in Lumen Document — one paragraph per flagged line
- Draft the board email with both files attached — needs your approval to send
- Read 1,284 rows · 6 lines moved more than 10% · 1 sign error flagged for you
- Built 14 slides — re-ran chart on slide 9, the axis had inherited Q2 scale
- Wrote 6 commentary paragraphs · 780 words · linked each to its source row
- Email drafted and waiting in your outbox — not sent
What you get back is not a transcript. It is a .lslides deck you can open and edit, a document with real paragraphs, and an email sitting in your outbox with the attachments already on it. The sign error it flagged is a link — click it and you land on the exact cell.
What it can reach
| Source | What it can do | Your control |
|---|---|---|
| Your library | Read and edit every Lumen file | On by default |
| Files on your computer | Read folders you point it at | You choose the folders |
| The web | Search and read pages for research | On by default, per-job |
| Connected services | Gmail, Google Drive, Calendar, Outlook, Slack | Off until you connect each one |
| Your screen | See what you are looking at, when you ask | Asks every time |
| Your own tools | Anything you expose over MCP | You define the tools |
Why the plan is written down
A plan you can read is a plan you can correct. In practice most corrections happen before the run, which is the cheapest possible place for them — you delete a step, change “last quarter” to “last two quarters”, and the whole job comes out right the first time. It also means you can hand the same job to Lumen next quarter and know precisely what will happen.
- Say the outcome, not the steps. “A board pack from these numbers” beats “open Slides, then…”.
- Name your sources. Mentioning the file or folder saves it a search and makes the plan concrete.
- Correct the plan, not the output. One edit before the run replaces ten edits after it.
- Let it check itself. On longer jobs it re-reads its own work; that pass is where most errors die.
- Set it running on a schedule. Anything you approve once can become a weekly automation.
From one job to a standing one
Once a job comes out right, you can put it on a trigger — a schedule, a new file in a folder, or an incoming message that matches a rule. The plan stays visible, the approval gate stays where it is, and the work simply arrives already done. Watching for a trigger costs nothing; you only spend credits when one actually fires.
Common questions about the agent
Can it do something I did not ask for?
It cannot act outside the plan you approved. If a job turns out to need a step that was not in the plan — a new file, a connector it has not been given — it comes back and asks rather than improvising.
What stops it emailing the wrong people?
Sending is never automatic. Messages are drafted with the recipients visible and held until you press send, including for jobs running on a schedule.
Does it need my files in the cloud?
No. It reads them locally and sends only what a given step needs for AI processing. Connected services like Gmail or Drive are separate, individually granted, and revocable at any time.
How much does a job like the board pack cost?
The example above ran to 19 credits. The estimate is always shown before you approve, and long jobs report their spend as they go so nothing surprises you at the end.
What if it gets something wrong?
Every file it touches has full version history, so any change can be rolled back. Its own verification pass catches a good share of errors first, and anything it is unsure about is flagged in the run log rather than quietly assumed.
