Agentic AI Capability

State what to watch in plain language; Juno watches it.

A watch is a rule that describes a condition you want to hear about. You write it the way you would say it, Juno turns it into a rule the engine evaluates as often as you choose, and the finding reaches the person who can act on it.

WHAT JUNO WATCHESONE WATCH, FOUR THINGS IT DOESWHERE THE FINDING GOESForecast and demanderror against whatactually happenedStock and coverweeks of cover againstyour floorOpen purchase ordersdates, quantities, windowsSupplier behaviourconfirmed dates againstreceiptsException messageswhat your planning runreportsThe plannera task with its evidenceThe buyera draft to approveA managera digest and the countsA channelTeams, Slack or a mailboxAn agentthe watch starts its run01Evaluate the conditionthe published rule, on the factsof the moment02Compare with your linethe threshold the watch carries03Record the findingfacts, rule version, value atstake04Route the alertthe role, the channel, the clockevery watch reads the same fact base an agent run reads, under the same access
One watch does four things, and the finding leaves by the route you set for it. The dashed arrow returns a finding that starts an agent run. Product.
12
rules under watch in the showcase estate
showcase estate
4
steps in the method: state, alert, resolve, execute
method
4
ways a watch fires: a schedule, an interval, an event, a person
product
10%
the forecast error that starts the showcase watch
showcase estate
Where this sits

Stage 5 of six: act

Step by step

From a sentence about what matters to an agreed resolution

Four steps: state the rule, alert the right person, resolve inside your limits, execute the agreed resolution.

FOUR STEPS, FROM A STATED RULE TO AN AGREED RESOLUTION01State the rule to watchin plain language, with the linethat matters to you02Alert the right personthe role that can act, on thechannel they already use03Resolve inside thelimitsinside your limits the agent acts;beyond them it engages a person04Execute the agreedresolutionthe action goes out, and theoutcome is checked against theobjectivethe manager report counts what the agent corrected and what a person corrected
The same four steps carry every watch, whichever signal it follows. Method.

A watch starts as a sentence. A planner writes “tell me when the forecast error for a material goes above 10% in a month”, and the Planner Agent turns it into a rule with a field, a test and the line you set.

You then choose how often Juno looks: at a set time, at an interval, when a row changes in a source, or when a person runs it. The watch reads the same fact base an agent run reads, under the same access.

When the condition is met, Juno records a finding: the facts at that moment, the rule version, the measure and the value at stake. The finding carries its reason in plain words, so the person who receives it can judge it in seconds.

The alert then follows the route the watch carries: a task in a queue, a message in a channel, a line in a digest, or the start of an agent run. Every route keeps the same evidence with it.

Under watch

Twelve rules, each with its own line, level and route

One screen holds every watch in the estate, with its line, its level, its route and what the last run found.

Rules under watch · showcase estate · 7 of 12
Rule under watchWhat it readsHow oftenLevelLast runWhat the last run found
Forecast error above 10%Forecast error by material and monthMonthlyA101 Sep 06:00Watching
Cover below the policy floorWeeks of cover per material at plantEvery 15 minA2Mon 07:053 findings
Cancellation window closingOpen lines and the window on each oneDaily 06:00A2Mon 06:0038 findings
Excess above eight weeksStock on hand and on order against demandDaily 05:30A2Mon 05:3014 findings
Supplier late rate above 5%Confirmed dates against receipt datesWeeklyA1Sun 22:00Watching
Lead time drift above 4 daysPlanned delivery time against actualWeeklyA1Sun 22:102 findings
An allocation-bound supplier changesAn event on the supplier recordOn the eventA2Fri 16:20Paused by P. Sharma

Seven of the twelve rules under watch. Showcase estate.

Each watch carries an owner, a level, a route and a version history. The owner edits the line, changes the route, pauses the watch or retires it, and each of those changes becomes a new version in the rule store.

The level decides what a finding may lead to. At A1 the watch reports and a person acts. At A2 it drafts an action and a named person approves it. At A3 it acts inside your limits and sends the rest to a person.

Where the alert goes

Five routes, each carrying the same evidence

The route is part of the watch, so the finding reaches the desk that can act on it.

RouteWhat arrivesWho usually receives itWhen it suits
A task in an approval queueThe drafted action, its evidence and a clockThe buyer or planner who owns that decisionA finding that needs a decision today
A message in Teams or SlackOne line with the value at stake and a link to the findingThe desk that watches the channelA finding people act on within the hour
A digest by e-mailThe findings of the period, grouped and rankedManagers and the planning leadA weekly or monthly review
An answer in ChatLimaThe finding with the terms, the systems and the query behind itAnyone asking about that material or supplierA question that follows an alert
An agent runThe finding starts the agent that owns that decisionThe agent, at the level you set for itA case your published rules already cover

A route is a setting on the watch, so the same finding can reach two places: a message for the person on duty and a task in the queue for the record. Each copy carries the finding reference, so a reply in one place is visible in the other.

Findings are grouped before they are sent. One material with 14 lines above the cover floor arrives as one alert with 14 lines inside it, ranked by value at stake. A clock on the queue and an escalation route to the next desk are set on the approvals page.

A watch at work

The forecast-error watch, written once and running monthly

One sentence, five rows, a monthly run, and a proposal a planner can accept or set aside.

Forecast error watchA1 Assistedshowcase estate
Watch the forecast error for each material and month. When the error goes above 10%, alert the demand planner. Where another method holds the error below 10% across the last six months, propose that method for approval.
What it compiles to: the planning table
RowFieldTestValueThen
01Forecast errorabove10% in a monthcarry on
02Material classequalsA or Bcarry on
03History heldat least6 monthscarry on
04Alertgoes tothe demand plannersend the alert
05Other methodstested onthe last 6 monthspropose the better method
Watch version 2Owner: L. TanMonthly, 06:00 on day 1Route: Teams and the planning queue

The sentence is the whole watch; the table below it is what the engine evaluates. Showcase estate.

1
The monthly watch runsforecast error is read for every material and month in scope
06:00
2
Twelve materials pass the lineerror above 10%, with 7741-X at 14.2%
12 findings
3
The planner is alertedone message carrying the value at stake and the working
L. Tan
4
Juno tests other methodsthree forecast methods replayed on the last six months
best 7.8%
5
A proposal is draftedthe method that held the error lowest, with the evidence beside it
A1 informs
6
L. Tan applies the methodthe planning parameter is updated and the watch carries on
applied

This watch sits at A1 assisted: it reports, and L. Tan decides. Juno still does the work of testing the alternatives, so the message arrives with a proposal and the evidence behind it.

Raising the watch to A2 would send the same proposal as a draft for a named approver. Raising it to A3 would let it apply a method inside your limits and send the rest to a person.

The limits are what make that safe. A method change applies where the tested error stays below your line across six months of history, and where the material sits in the classes the watch covers.

Every firing of the watch is on the audit trail: the facts read, the version evaluated, the alert sent, the person who acted and what happened afterwards.

The running example

Monday 05:30 to 07:40: three watches, one ranked morning

Showcase estate.

The running example

The excess watch ran at 05:30 and found 14 materials above eight weeks of cover. At 06:00 the cancellation window watch found 38 open lines worth $412,000 whose windows close on Thursday. Both findings carried the material, the plant, the value at stake and the working.

By the time L. Tan opened the 1,240 exception messages at 07:40, both findings sat at the top of her list: the 212 messages that touch an open purchase order line, with the 38 cancellable lines marked. A shortage triage role, once built from the blueprint, would rank the rest of the morning the same way.

Next capability

Recommendations with their basis

What a finding becomes when the evidence supports an action: a recommendation with its basis, its confidence and the value at stake.

Terms on this page
WatchA published rule that describes a condition you want to hear about, with the line that matters.
LineThe threshold a watch compares against, such as a forecast error above 10%.
FindingA recorded match between a watch and the facts at that moment, with its reason.
RouteWhere a finding is sent: a queue, a channel, a digest, an answer or an agent run.
DigestA grouped summary of the findings of a period, ranked and sent to a named role.
Manager reportThe count of corrections the agents made and the corrections people made, per agent and per period.
Questions
Who decides which alerts we receive?

The owner of each watch sets the route and the level. A manager can add a digest for the team, and anyone can ask ChatLima for the findings of a watch at any time.

How do we keep alerts down to a useful number?

Findings are grouped by material, supplier or order, ranked by value at stake, and sent once per period per line. Raising the line on a watch is a single edit, held as a new version.

Can a watch start an agent?

Yes. A finding is one of the four ways an agent run begins, so a watch on stock cover could start a stock rebalance role at the level it carries, once that role is built from the blueprint.

Can we watch a measure of our own?

Yes. Add the measure to the ontology, map the columns that feed it, and write the watch against it in the same plain language.

In the platform

Where Juno sits among the four capabilities

Each one builds on the one before it.

One shared foundationCatalogLineageAccess policyWrite path

See Juno answer and act on your own data

Bring two systems and one decision your planners make every week. In the first week we register the data, confirm the mapping with your team, answer your questions in ChatLima and run the first agent at the level you choose.