Desk agent widget

A Desk agent is an AI worker that lives on a desk. You wire other widgets into it, such as notes, a table or a browser, and give it a standing instruction, for example "keep a running summary of the wired notes". It reads what is wired in and does that job when you click Run, whenever one of its inputs changes, or on a regular interval. Its result shows in the widget, and it can write the result into a note for you.

Use it for work you would otherwise repeat by hand: summarising notes as they grow, checking a few web pages for anything new, filling in missing details in a table, or drafting from the material on a desk.

More about this widget Coming soon

Before you start

• The Desk agent is on every plan.
• AI must be set up. When you are signed in, agents run on your PlexiDesk credits. You can also use your own Anthropic key. See AI credits and where your AI runs Coming soon and Add your own API keys.
• You need a desk open. An agent reads widgets on its own desk. To bring in another desk, wire in a Portal widget Coming soon that shows it.

Add a Desk agent

1. Open the desk you want the agent on.
2. Click Add widget in the centre of the header.
3. Click Desk agent.

The Add widget menu groups widgets into columns by what they are for. Desk agent is in the Build & automate column. You can also type desk agent in the search box at the top of the menu and press Enter.

You can also press ⌘ K on a Mac or Ctrl K on Windows, type desk agent and choose Add Desk agent.

Start from the widgets it should work on

• From one widget. Right-click a widget's title bar and choose Create & link > Automate with an agent. A new agent appears beside the widget, with the widget already wired in.
• From several widgets. Hold Shift and click the title bar of each widget to select them. Right-click the title bar of one of them and choose Organise > Automate 3 with an agent (the number matches your selection). One agent appears to the right of them, with all of them wired in.

Ask Plexii to set one up

Tell Plexii about a job that should keep happening, for example "set up an agent that keeps a running summary of my meeting notes". Plexii proposes the agent, and can propose the wires into it, as cards for you to review. When you apply the card, the agent appears on the desk with Active unticked, so it does nothing until you tick Active or click Run. See Review and apply Plexii's changes Coming soon.

What is in the widget

From top to bottom:

• The role the agent works in. It starts as Generalist.
• Quick jobs, shown while the instruction is empty.
• The instruction box, for the agent's standing instruction.
• The inputs line, which counts the widgets wired in and names the first three.
• Manual, On change and Interval: when the agent runs.
• Write to: where the agent's result goes, besides the widget itself.
• Active and Run, with the interval in seconds when the agent runs on an interval.
• The result, and any Proposed changes for you to review.
• The footer: when the agent last ran, and how many runs are logged.

Wire in its inputs

An agent reads only the widgets wired into it.

1. Point at the widget you want the agent to read, then click the link button in its title bar. Its tooltip says Click, then click another widget to connect them.
2. Click the agent.
3. In How should this connect?, choose Feed in as context.

Repeat for each widget you want the agent to read. The inputs line changes from No inputs — wire widgets into this agent to the count and names, for example 2 inputs: Meeting notes, Budget. See Link widgets and set up wires.

What an agent can read

• Text widgets such as a Sticky, Note, Markdown, Page or Card: what is written in them.
• A Table: its rows.
• A Field or a Custom block: each label and value.
• A File or link widget: the text inside a PDF, a Word .docx file, a spreadsheet or a text file.
• A Browser: the page it is showing, including pages you are signed in to. The agent can also use the browser to research. See Research the web with a Browser.
• A Portal: the whole desk the portal shows.
• Another Desk agent: its instruction and its latest result.

From each input, the agent reads up to the first 2,000 characters. For a long document, wire in the part you need, or a summary of it. When the agent researches with a Browser, it also reads the pages it opens there.

Give it a role

The role shapes how the agent approaches the work: what it pays attention to, its judgement and its tone. Rest the pointer on the role to read its full description.

1. Click the role at the top of the widget.
2. Click a role, or type in Search roles… to find one.

The list starts with six general roles:

More specialist roles follow them in the list.

Create your own role

1. Click the role, then Create new profile… at the bottom of the list.
2. In New agent profile, describe what the agent should be good at, then click Draft with AI to fill in the rest. Or fill it in yourself.
3. Check the Profile name, the One-line summary and How it works (its expertise + approach), and pick an icon.
4. Click Save profile. The agent takes the new role.

A name and How it works are both needed. Roles you create are saved on this computer and appear in the role list of every agent on it. To delete one, point at it in the list and click its bin. See Choose or create an agent role Coming soon.

Write its instruction

Start from a Quick job

While the instruction box is empty, Quick jobs above it offer five common jobs. Each one fills in the instruction, picks a role and sets when the agent runs. You can change any of them afterwards.

Write your own

Click in the box, which suggests Standing instruction, e.g. keep a running summary of the wired notes, and say what the agent should do. Your instruction is saved as you type.

• Give it one job. "List the decisions in the wired notes, one per line" works better than an instruction that asks for four things.
• Say what to produce, not just what to look at.
• Type @ to pick a desk, widget, document or person by name. Naming something does not wire it in: the agent reads only the widgets wired into it.

Suggestions while you type

• Better fit. When another role clearly suits your instruction, a strip appears under the box: Better fit: and the role's name. Click Switch to give this agent that role, or + linked to add a second agent with that role beside this one, with the same instruction and settings, wired to the same widgets. Click × to hide the suggestion. This suggestion is worked out on your computer, without AI.
• Design a specialist for this. Once your instruction is more than ten characters long, this link appears under it. Click it and AI writes a new role made for your instruction (the link shows Designing… while it works). A card shows the new role's name and summary. Click Replace this agent to give this agent the new role, or Create linked agent to add a second agent with it, wired to the same widgets. Either way, the role is saved to your role list. Click × to dismiss it.

Choose when it runs

Click one of the three buttons:

Active is ticked on a new agent. Untick it to stop On change and Interval runs. Run still runs the agent once whenever you click it.

Run it

1. Click Run. The button shows Running…, and the widget shows Thinking… until the first result arrives.
2. Read the result in the widget.

The footer shows when the agent last ran, for example ran 5m ago, and how many runs are logged. The agent keeps a log of its last eight runs.

If something stops a run, the reason shows in red where the result would be. For example, clicking Run before writing an instruction shows Give the agent an instruction first.

Where its result goes

The result always shows in the agent. Each new run replaces it. To keep results somewhere you can read and build on, write them into a note, or wire the agent to another widget.

Write each result into a note

1. Open the Write to list. It starts at Nowhere — just show the result.
2. Choose a Note or Markdown note on this desk, listed by type and name. Or choose Add a Note… or Add a Markdown note… to add a new one to the right of the agent, named after it, for example Desk agent output.
3. Choose how each run writes there. The button beside the list switches between append, which adds each result under what is already there, and replace, which swaps it. A new agent starts on append.
4. Run the agent. The result arrives as a card under Proposed changes in the agent. Click the card to write it in, or click × on it to dismiss it.

To skip the card for this destination, tick Write without asking me first. Each run then writes straight in, and the result in the agent starts with Written to and the note's name. Anything else the agent proposes still arrives as a card. Choose Nowhere — just show the result to stop writing anywhere.

Keep a Page or a Table up to date with a wire

Draw a wire from the agent to the widget, and choose Keep in sync when asked how it should connect. After each run, the agent's latest result is written into that widget in its own form, without a card:

• A Page, Note, Markdown note or Sticky is replaced with the latest result. A Page gets real headings, bullet lists and paragraphs.
• A Table is updated by the table's AI: rows that match are updated, new items are added as rows, and other rows are left alone.

The inputs line then also shows feeds and the number of widgets the agent delivers to. To see what a wire wrote, or undo a write with Revert this write, click the line. See Link widgets and set up wires.

Proposed changes

Besides its result, an agent can propose changes to the widgets wired to it, such as filling in cells or adding rows in a wired table, or adding to a wired document. It can also propose new things, such as a document, a desk or a draft email. Each one arrives as a card under Proposed changes, saying what it would do and why.

• Click a card to apply it.
• Click × on a card to dismiss it.
• When there are several, tick the boxes beside the ones you want and click Apply selected, or click Apply all to apply every card. Each button shows how many cards it will apply.

Nothing is changed until you apply a card. A new run replaces any cards you have not dealt with.

Research the web with a Browser

Wire a Browser widget into the agent and the agent can use it. It reads the page that is showing, can open other pages and can search the web in that browser, then writes up what it found.

• The Browser goes to the pages the agent visits, so give the agent a Browser of its own rather than one you are reading.
• With a Browser wired in, the agent researches and reports. It does not propose other changes, but Write to and wires from the agent still deliver its result.

See and manage your agents

• On this desk: the Automations panel lists every agent and wire on the desk, when each last ran and whether it worked, with a switch to turn each one off. See See what runs on its own on a desk.
• On every desk: in Brain, open Agents to see every desk agent in your workspace, the desk it is on, whether it is active and when it last ran. Click an agent to go to its desk. See See and manage all your agents Coming soon.

When you publish a live view of a desk, a Desk agent shows as a placeholder. Its instruction and results are not published. See Publish a live view of a desk.

Good uses

• A running summary. A Summarise agent wired to your meeting notes, writing to a Note. See Keep a running summary with a desk agent Coming soon.
• A watch on a web page. A Monitor agent wired to a Browser on a page you check often.
• A fuller table. An Enrich a table agent wired to a table of companies, proposing the missing details for you to apply.
• A second pair of eyes. A Critic / QA agent wired to a draft Page, run before you send the draft.
• A research digest. A Research agent wired to a Browser and to a Note listing what to look into, writing its findings to another Note.

Common questions

My agent does not run on its own

Check that Active is ticked and that the trigger is On change or Interval. An On change agent runs when a wired input changes, so check the inputs line for the widgets you expect. An Interval agent runs while its desk is open.

The result shows in the agent but not in my note

Check that Write to names the note. Unless Write without asking me first is ticked, each result waits as a card under Proposed changes until you click it. If the agent says it described its work instead of writing it, nothing was saved: change the instruction to ask for the finished text itself, for example "Write the summary", and run it again.

Does an agent change my widgets without asking?

Only in two cases you set up yourself: a Write to destination with Write without asking me first ticked, and a wire from the agent to another widget. Everything else it proposes waits for you to apply it.

Can an agent read a widget on another desk?

It reads widgets on its own desk. To bring in another desk, add a Portal widget Coming soon that shows it, and wire the Portal into the agent.

Are the roles I create on my other computers?

Roles you create are saved on the computer where you made them.

The agent shows a message about AI or an API key

The agent needs AI to run. Sign in to use your PlexiDesk credits, or add your own Anthropic key. See AI settings.

Related tasks

• Build & automate widgets
• Choose or create an agent role Coming soon
• See and manage all your agents Coming soon
• Link widgets and set up wires
• See what runs on its own on a desk