Running Agentum from Chat
Most of what you’d click through in Agentum, you can just ask for. Chat isn’t only where your agents talk to you; it’s a way to run the app. Type what you want in the ✨ Agentum thread — on your Mac, or from your phone in Web Chat — and the assistant does the work.
You never need to know an agent’s technical id. Use whatever you call it: “the weather one”, “My Network”, “news”. If it’s ambiguous, you’ll be asked which one you meant, by name.
The one rule: nothing changes without your tap
Anything that only reads happens immediately — “what’s my news agent set to?” just answers.
Anything that changes something first posts a card in the chat: a plain-English description of what’s about to happen, a button that does it (labelled for the job — Start, Install, Update, Confirm), and Not now. Nothing has happened until you tap. If the assistant ever says “I’ve done it” while a card is still sitting there untapped, it’s wrong — the card is the truth.
Change how an agent behaves
This is the big one, and it’s new. Say what you want different:
“Have the weather agent include wind speed.”
“Stop the news agent sending on weekends.”
“Make my morning digest shorter and send it at 6:45 instead.”
You get a card naming the agent and quoting back the change. Tap Start, and a real coding assistant begins working on your Mac — the same one the Improve tab uses, just started from a sentence instead of a form.
Three things the card tells you up front, because they’re worth knowing before you tap:
- A restore point is saved first, so the change can always be rolled back.
- It only runs while Agentum is open on your Mac. There’s no server doing this for you; quit the app and the session ends.
- Where questions will reach you. If your coding assistant supports it, it asks you right here in the chat and you answer in ordinary words — your reply gets passed straight through to it. If it doesn’t, the card says so, and the question waits for you in the app instead.
While it’s running you can ask “is it done yet?” at any time, or tell it to stop. When it finishes, it says so in this same conversation.
Two limits worth knowing: this works on agents you own, not Marketplace ones (those are sealed read-only — change their behaviour in their Settings tab instead), and it’s macOS only.
Build a brand-new agent
“I want something that chases my landlord about the deposit.”
The build walks you through it as a series of cards right in the chat — a few questions, a plan to confirm, then it builds. It’s the same flow as the Create Agent page (Creating Agents), available wherever you happen to be talking.
Find and install agents
“Is there an agent for tracking expenses?”
“Any updates for my agents?”
Searching the marketplace is a read, so you get the answer straight away. Installing and updating are changes, so they arrive as a card to approve — including Update all. Your assistant may also raise this on its own if it notices you keep asking for something an agent already does well.
Ask what an agent is set to, and change it
“When does my reminders agent start in the morning?”
“What can I change on News?”
“Set the digest time to 8:15.”
Reading settings and reading an agent’s own data (“what’s in my reading list?”) answer immediately. Changing a setting is a card.
Turn agents on and off, and manage what they can reach
“Turn off the worklog agent for now.”
“Does the news agent have access to my calendar?”
“Stop the travel agent from reading my email.”
Switching an agent off and changing what it’s allowed to touch both take effect on the running system straight away once you approve — no restart.
Have an agent do something now
Agents publish their own actions, and you can trigger them by asking:
“Send the news digest now.”
“Add milk to the shopping list.”
Harmless actions run straight away. Anything with a real consequence — sending, spending, deleting, anything that reaches outside the app — asks you to confirm first.
Walk through an agent’s setup
“Help me set up the calendar agent.”
Agents that ship a setup guide can run it as a card that stays put and fills in as you go, step by step, without you hunting through settings screens.
Set up something recurring
“Every weekday at 7, give me my calendar and anything that came in overnight.”
That creates a routine — a saved prompt your assistant runs on a schedule. You can see, edit, pause, and delete all of them on the Routines page in the sidebar. Ask in plain English to change one; there’s no syntax to learn.
Quieten an agent without switching it off
“Snooze worklog for a week.”
A snoozed agent keeps running and keeps its data — it just stops messaging you unprompted for a while. It still answers when you ask it something.
When something can’t be done
The assistant is meant to tell you plainly rather than pretend. Common ones:
| You’ll hear | What it means |
|---|---|
| “Agentum isn’t running” | Coding sessions and installs only happen while the app is open on your Mac. |
| “That one’s a Marketplace agent” | Its code is sealed. Change what it does in its Settings tab instead. |
| “It’s already being worked on” | One coding session per agent at a time. Ask for its status, or wait. |
| “It’s waiting on you at the Mac” | A session is stuck on a question or a permission prompt. Go tap it there. |
Where to next
- Improving Agents — the full-screen version of the same thing, with a diff to review
- Creating Agents — the guided build flow
- Agentum Web Chat — all of the above, from your phone