Turning On an Agent and Reading Its Runs

An agent in the library is a template. Turning it on creates your own copy of it, with your settings, your alerts and your schedule.

When you’d use this

You found an agent in the library and want it working on your data.
You want the same agent twice, set up differently for two departments.
An agent raised something and you want to see what it actually read.
An agent should have run and did not, and you need to know why.

Finding it

Sidebar → Agents → Agent library. Click Set up on a library card to create a copy. Once it is on, the agent appears under Sidebar → Agents → My agents and its card links to its own page.

Owners and admins can always open Agents. Anyone else needs a role with agent access, such as Agent Operator: it lets a person see the agents, build one from a sentence, turn one on, pause it and change its settings. An agent they build can use only the records they can read themselves. If you can view agents but not change them, the settings read “You can view these settings. To change them, ask someone with permission to edit agents.”

Setting one up

Setup is a short wizard that writes nothing until the last step, so you can leave at any point and nothing is turned on. Leaving mid-way confirms first: “Leave setup? Your choices are not saved and nothing is turned on.”

Which steps you see depends on the agent:

Settings - the values this agent exposes. Only shown when it has any you can change.
Alerts - always shown. Choose whether findings raise inbox tasks.
Schedule - only when the agent ships with one. It is read-only here: “This agent runs on the schedule it ships with. You can change it on the agent’s Schedule tab once it is on.”
Test with these settings - only when the agent has a test profile. Running a test uses the same AI as a real run, and nothing it does is saved. The result tells you plainly whether it found nothing, or how many inbox tasks it would raise.
Review and turn on - name the copy, check the values, and click Turn on.

Only the settings you actually changed are saved, so the agent keeps following the library default for everything else - and picks up an improved default later.

The agent’s own page

Once on, the agent has eight tabs:

Tab	What it holds
Overview	How it is doing: hours saved, items caught, runs that finished, and when it runs next
Settings	This copy’s values, with the history of who changed what. Also its model, monthly limit, Always ask me when… rules, what it can read for reference, reviews of real runs and who hears about problems
Alerts	Whether findings raise inbox tasks, and to whom. Also Who approves this agent’s changes: the owner, a backup approver and how long changes wait
Schedule	When it runs on its own
Runs	Every run, newest first
Evaluate	Checks built from past runs, and comparisons of a change against how it works today. See Evaluate an agent before you change it
History	Changes to this copy over time
Version	The published definition version it runs

Header controls: Run now, Turn on / Pause, and Make a copy. A copy arrives turned off (“This copy is turned off. Review its settings, then turn it on.”) so you can retune it before it does anything.

Run now is disabled while the agent still needs setup, and says so: “Finish setting up this agent before it can run.”

Schedule

Open the Schedule tab and pick from Schedule:

Not scheduled - runs when something triggers it
Every hour
Every day at 6:00am
Every Monday at 6:00am
The 1st of every month, 6:00am
Custom schedule…, which gives you a Cron expression field (for example 0 6 * * 1-5 for 6am on weekdays)

Click Save schedule. The card then shows the schedule as a badge, or Not scheduled.

Under When something happens you can also run the agent on an event: pick it from Run this agent when, narrow it with a condition or leave it at Every time, and click Add trigger.

Saved settings sets

If several copies should share one configuration, save the values as a set and point the others at it. The page confirms with “This agent now follows those saved settings”, and Change reverted if you back a change out.

How it is doing: the Overview tab

The Overview tab opens first. It shows what this agent did for you over the last days:

Hours saved and Items caught.
Runs that finished, Tool errors, and how often You approved and You rejected its findings, each with the change from last week.
Credits per run and Credits this month, against its monthly limit when one is set.
When it runs next, with Run now and Open Inbox.

Before the first run, the tab lists what happens next instead of empty numbers.

How the hours saved number is worked out

Each result an agent produces carries an estimate of the time it saves you: the usual minutes to do that kind of task by hand, minus the minutes it takes with the agent. A result counts only once it is realized, meaning you approved it or acted on it. A finding still waiting for you, or one you dismissed, does not count. Hours saved is the total of those minutes for the last 30 days, shown in hours to one decimal place. Items caught counts the same realized results. A custom agent you wrote from your own instructions has no standard time for its work until one of its results declares one, so until then its Hours saved tile reads Not measured for custom agents instead of a number. A copy of a built-in agent is measured like the built-in.

Sidebar → Agents → Agent health shows the same picture for every agent you turned on, with the ones that need your attention first.

Limits and rules

These sections are on the Settings tab. Only the agent’s owner or a workspace admin can change them.

Model

How careful this agent is sets the model: Fast, Balanced or Smart. Fast is cheapest, Smart is most careful.

Monthly limit

Monthly limit (credits) pauses the agent when it reaches that many credits in a month. Leave it empty for no limit. The section shows what it used this month, for example “Used 120 of 400 this month”.

Always ask me when…

Write, in plain words, when the agent should stop and check with you, for example “the amount is over $1,000”. Click Preview to see the rule Muin understood, then Save rule. If Muin cannot turn your words into a rule, it says so and offers examples.

These rules only add questions. Changes the agent proposes always ask you anyway, and a payment or an invoice approval always needs a person to approve it. Only a person creates or changes a rule; an agent cannot. Edit puts a rule’s words back in the box: change them, click Preview again, then Save changes (or Cancel to leave the rule as it was). An edit is checked the same way as a new rule. Remove deletes a rule after you confirm.

What it can read for reference

Pick documents, folders or websites the agent may read while it works. A finding that used one shows “From:” and the source’s name. The agent’s owner must be able to read every source you pick.

Review real runs

For an agent you built, Review real runs has a careful model check 5% of its finished runs against the agent’s instructions and every tool it used. It is off until you turn it on, and the reviews are charged as AI credits. The section then shows how many runs were reviewed and how many issues were caught.

Who hears about problems

The agent’s owner, plus anyone you add under Also tell, is told when:

A run fails, if Email when a run fails is on.
Its findings start getting rejected more often than last week, or its tool errors rise. With Review real runs on, also when fewer reviewed runs pass than last week. This notice is sent at most once a week.

Telling it when it is wrong

Each finding and each run has thumbs up and thumbs down under “Was this helpful?”. With thumbs down you can say What should it have done? When you dismiss a finding, pick why: Not a problem, Already handled, Wrong record or Other. Your feedback is stored with the agent and the run it is about.

Going back to an earlier version

For an agent you built, every saved prompt is kept as a version. Open the agent from Sidebar → Agents → My agents → Custom, then Versions, and click Roll back on the one you want.

Reading a run

Open the Runs tab and click a run, or follow a link from a finding.

A run is in one of seven states: pending, running, waiting, completed, failed, cancelled or timeout. While it is pending or running, the page refreshes itself every few seconds, and offers Cancel. A failed run offers Retry.

A run is waiting when the agent proposed a change to your records. Agents never change anything until a person approves: the run page reads “Waiting for Omar Siddiqui to approve the changes this run proposed.” with an Open in Inbox link. When the change is decided, the run ends with the outcome in plain words. A run that did not make its change is marked cancelled and shows the reason:

Approved: the change runs and the run finishes with the result.
Rejected: “Rejected by Omar Siddiqui: “Wrong claim”. Nothing was changed.”
Expired: “Expired: nobody approved within 7 days. Nothing was changed.” with Run again.
Stopped: “Stopped: Omar Siddiqui no longer has access to this agent. An admin can turn it back on.” Its waiting changes are withdrawn and the workspace admins are told.

See Approving Agent Changes.

The run page shows:

Findings - what the agent concluded, and the inbox tasks it raised.
Extracted Fields - the values it read out of the document.
Execution Steps and AI Decisions - the sequence it worked through, each marked Completed, Failed, Skipped or Running.
Quality Flags - anything the agent was unsure about.
Raw output - the unformatted result, behind a disclosure.
Error - the code, the message and the detail, when it failed.

Along the side: the trigger and triggering event, the definition version, start and finish times, duration, confidence, the model used, input/output/total tokens, retry count, and links to the document and the agent copy.

An agent that worked purely from your records rather than a document reports “Not used. Worked from your records.” instead of a model and token count. Runs record tokens, not a dollar figure.

Two notices worth knowing

A run queued for more than about a minute and a half shows Not started yet, and states that nothing has run and nothing has been charged.
A run whose document is missing shows No document to read and says retrying will not help. Pick a document and run again.

What it affects

Turning an agent on means it runs on its schedule, on your live data, without further prompting. It reads freely, but every change it proposes to a record waits for its owner’s approval in Inbox, unless a standing rule the owner set covers it. Payments and invoice approvals always wait.
The agent acts with the access of the person who turned it on, never more.
With alerts on, findings become real inbox tasks for the role you chose. See Tasks.
Pausing stops future scheduled runs. It does not undo anything already done.
Each run uses AI credits. A monthly limit pauses the agent before it uses more. See AI Credits.
Changing a copy’s settings affects only that copy. Changing a saved settings set affects every copy following it.

When it goes wrong

What you see	Why	What to do
“This agent could not be loaded”	The agent is no longer in the library, or the request failed	Click Back to the library and pick it again
“You do not have permission to set up agents.”	Your role cannot create agent copies	Ask an owner or admin
“The test could not finish: … Nothing was saved.”	The test run failed	Click Try again. Your choices are still there
“Could not reach Muin. Your choices are still here.”	The final save did not land	Try Turn on again. Nothing was created twice - the setup carries a key that prevents duplicates
A validation error sends you back to Settings	One of your values was out of range or the wrong type	Fix the field it points at and continue
“Could not start the run” / “Could not change this agent’s status”	The request was refused	Check the agent is not already running, then retry
“This agent has no activation yet, so it cannot be scheduled.”	You are looking at a definition with no copy turned on	Set it up first
“Run not found”	The run was from another organization, or no longer exists	Click Back to Run History
A run sits in pending	Nothing has picked it up yet	Wait. The page tells you once it has been queued too long, and confirms nothing was charged
A run sits in waiting	It proposed a change and is waiting for a person to approve it	Click Open in Inbox and approve or reject the change

Next step

Building a Custom Agent - when nothing in the library fits.
AI Agents Overview - what each library agent does.
Task Management - where an agent’s findings land.