Agentic AI Chapter 4 of 9

Everything a person does with an agent is a setting or a signal.

At every level a person is in the picture. This chapter shows what that person does: the four patterns of working with an agent, the screens each person sees, the nine settings that shape every agent and who owns each, and the feedback loop in which every approve, edit and reject teaches the next cycle.

PEOPLELTL. Tanplanner · owner of the agentDOD. Okaforbuyer · named approverJHJ. Harrisonfinance lead · limitsJRJ. Ridleydirector · sets the levelTHE AGENTPO cancellation agentruns at 06:00 on version 338 lines match, worth $412,00029 drafts under $25,000 to thequeue9 lines above it to D. Okafor asbuyerwrite-back kept with itsacknowledgementone audit record per decisionA2approver · D. Okaforpolicy version 3 · released Fri 18 SepSETTINGSobjective · data scope · rules · trigger · actions · limits· escalation · levelWORKa draft with its basis and its confidence · an exceptionwith its reason · the audit record · the manager reportSIGNALSapprove · edit · reject · override, each recorded and readby the next cycle
Everything that crosses from the people to the agent is a setting or a signal; everything that crosses back is work with its reasons attached. Showcase estate.
4
patterns of working together, one for each level where an agent runs
product
9
configurable elements on every agent, each with a named owner
product
29
drafts worth $187,300 in D. Okafor’s approval queue on the Monday
showcase estate
203 of 212
drafts approved as written over six weeks; the other 9 were edited
showcase estate
Where the story stands
So far
  1. The ladderYou have seen the five levels and a worked scene at each: what the agent does, what the person does.
  2. What it isYou have seen the eight parts of an agent.
  3. Why nowYou have seen that lost corrections are one of the four reasons earlier AI stalled.
Working together

Four patterns, one for each level where an agent runs

What passes between the agent and the person is different at each level, and so is the screen.

The ladder gave each level a definition. Read from the person’s chair, the four levels where an agent runs are four patterns of working together, and each is named by what passes between the two. At A0 the work stays with people and ChatLima answers their questions, as J. Harrison’s credit hold showed in the last chapter.

A1 · Assisted

Inform

The agent informs; the person decides and acts. What passes: a ranked list with its evidence. As designed, the shortage agent ranks Monday’s shortages by revenue at risk and L. Tan decides who gets the parts.

A2 · Supervised

Draft-and-approve

The agent decides and drafts; a named person approves each action. What passes: a draft with its basis and its confidence, then the go-ahead. The PO cancellation agent drafts 29 cancellations and D. Okafor approves each one.

A3 · Delegated

Execute-and-escalate

The agent executes within versioned rules; the person handles exceptions only. What passes: the exceptions, each with its reason. As designed, the baseline forecasting agent refreshes stable SKUs inside its error band and L. Tan takes the SKUs that breach it.

A4 · Autonomous

Execute-and-audit

The agent executes and self-corrects; the person audits after the fact. What passes: the record of every action, read as a sample. As designed, the cash application agent matches, re-tries and re-routes, and J. Harrison audits a sample each week.

Three of the four scenes are designed roles from the agent blueprint, yet to be built. The A2 scene is the PO cancellation agent, which runs in production at a high-tech manufacturer. blueprintproduction

The draft-and-approve pattern in full: the PO cancellation agent on the Monday

The PO cancellation agent A2 Supervised
  • Runs at 06:00 on version 3 of the policy and finds the 38 open lines that match it, worth $412,000.
  • Drafts a cancellation for each of the 29 lines under $25,000, with its basis in the words of the policy and a confidence.
  • Lists the 9 lines of $25,000 and above for D. Okafor as buyer, for his judgement, with the reason for each.
  • After the go-ahead, writes each change to the ERP through its released interface and keeps the acknowledgement.
a draft with its basis and its confidence, then the go-ahead
D. Okafor · buyer · named approver
  • Opens the queue and reads each draft: the reason, the values the agent read, the confidence and the value at stake.
  • Approves as drafted, edits a quantity or a date, or rejects with a reason. The 29 decisions take 22 minutes; the last is at 09:12.
  • Answers the 9 queries above the threshold with his own judgement, and each answer is recorded the same way.
  • Leaves 29 recorded decisions with his name and the time, and the next cycle reads every one of them.

Read the two columns together and the handoff is the whole relationship: the agent hands over a draft it can explain, and the person hands back a decision with a name on it. The rest of this chapter is about what the person sees and sets on each side of that line.

The screens

One screen for each person, and each screen shows the reasons

The buyer approves, then handles exceptions only; the finance lead reads one record; the director reads the report.

Approvals · D. Okafor · PO cancellation agent · Monday 21 September 2026 · 29 drafts · $187,300 · 5 of 29 shown
LineWhat the agent drafted, and whyValueConfidenceDecision
4500067123 · 00020Cancel 200 of 7741-X at plant 1210
excess 11.4 weeks · window open to Thu 24 Sep · supplier class standard
$8,42091%Approved as drafted · 09:12
4500068004 · 00020Cancel 1,200 of 8802-K at plant 1210
excess 9.3 weeks · window open to Wed 23 Sep · supplier class standard
$24,10093%Approved as drafted · 08:58
4500067990 · 00030Cancel 300 of 5513-B at plant 1220
excess 8.6 weeks · window open to Thu 24 Sep · supplier class standard
$7,16088%Edited to 200 · approved 09:03
4500068110 · 00010Cancel 150 of 9021-M at plant 1220
excess 12.1 weeks · window open to Fri 25 Sep · supplier class standard
$5,90090%Approved as drafted · 09:05
4500068221 · 00050Cancel 80 of 4410-T at plant 1210
excess 10.2 weeks · window open to Thu 24 Sep · supplier class standard
$2,28090%Approved as drafted · 09:10

Each row carries the action, the basis in the words of the policy, the confidence and the value at stake. The three decisions are approve, edit and reject with a reason. Showcase estate.

Exceptions · PO cancellation agent · A3 for lines under $25,000 · a later Monday · 3 open
LineWhy it reached a personWhat the agent didFor the person
4500069310 · 00010Value $31,600, above the $25,000 thresholdDrafted the cancellation and sent it to D. Okafor as buyer, with the reasonBuyer · D. Okafor · approve · edit · reject
4500069402 · 00030Confidence 0.76, under the 0.80 floor: two systems report the supplier class differentlyPaused the run on this line, reported both readings and told the ownerOwner · L. Tan · choose which reading to trust
4500069188 · 00020The write-back returned an error: the line is blocked in SAPHeld the change and sent it to the owner with the messageOwner · L. Tan · clear the block

At A3 the queue holds only what a limit or a thin finding handed over, each with the reason. Everything else was executed inside the rules, written back and confirmed. Showcase estate.

Audit view · DEC-2026-09-21-0417 · opened by J. Harrison, finance lead
PartWhat the record holds
Signal06:00 run · exception message for 7741-X at plant 1210 · 38 open lines read from SAP S/4HANA and Oracle Fusion
ReasoningPolicy version 3 · three conditions matched: excess 11.4 weeks, window open to Thu 24 Sep, supplier class standard
Recommendation06:06 · cancel 200 of 7741-X, the open quantity on line 00020 of order 4500067123 · $8,420 · confidence 0.91
Limits checkedValue under $25,000 · daily budget 29 of 60 · confidence floor 0.80 · all passed
ApprovalD. Okafor · approved as drafted · 09:12, after 3 h 06 min in the queue · role and session recorded
ActionPurchase order change posted to SAP S/4HANA through its released interface
AcknowledgementMessage 20260921-114, stored as returned
OutcomeDay 14: projected cover 6.2 weeks against 11.4 · verdict met
Manager report · J. Ridley · PO cancellation agent · six weeks, 21 September to 30 October 2026
MeasureSix weeksPer 1,000 decisions
Drafts produced (decisions)2121,000
Approved as written203958
Edited, then approved942
Rejected00
Overridden00
Exceptions handed to a person
at A2, a line above the value threshold, sent to D. Okafor as buyer
66311
Write-backs confirmed by the ERP2121,000
Line the team set for a move to A395% approved as written, held four consecutive weeksmet · weeks 3 to 6

The audit view holds the eight parts of one decision, the record chapter 5 reads in full. The manager report counts one agent over six weeks; each agent built later from the blueprint would be counted the same way. An exception at A2 is a line above the value threshold, which reaches D. Okafor as buyer with the reason. Showcase estate.

Each screen is a SmartApp over the same decision records, cut for one person. The buyer’s queue ranks drafts by value and time left, with a clock per queue and a named next desk when the clock runs out. The auditor opens one record by its reference; the director reads counts per 1,000 decisions, so agents of different sizes would compare once more than one is built.

The screens differ; the record behind them is one. The draft D. Okafor approved at 09:12 is the record J. Harrison opens by its reference and one of the 212 decisions J. Ridley counts. That is what lets a level review rest on evidence.

The configurable elements

Nine settings shape every agent, and each one has an owner

Seven are the parts you met in chapter 2; the level and the feedback decide how much the agent may do and how it improves.

Seven of the nine you met in chapter 2 as parts of the agent: objective, data scope, rules, trigger, actions, limits and escalation. The eighth part, the tests, is how a change to any of them is proved before it takes effect: a simulation on last quarter for a rule, a test on recorded cases for a level. The level and the feedback turn the parts into a working relationship.

Agent settings · PO cancellation agent · version 3 · A2 · showcase estate
ElementSettingSet byChanges
ObjectiveCancel purchase order lines that stand above the need, inside their cancellation windowThe planner, in a sentenceNew version
Data scopePurchase order lines, schedule lines, demand, stock and supplier class at plants 1210 and 1220The data leadNew version
RulesThe policy sentence and its planning table, version 3The planner with the Planner AgentNew version, simulated first
Trigger06:00 each working day, after the planning runThe plannerAt once
Actions permittedDraft a cancellation · send a change · open a query to the buyerThe ownerNew version
LimitsValue threshold $25,000 · daily action budget 60 · confidence floor 0.80L. Tan and J. Harrison, financeNew version
EscalationAbove the threshold, to D. Okafor as buyer, with the reason · a failed write-back, to the owner, L. Tan · a limit breach pauses the run and tells the ownerThe ownerAt once
LevelA2 Supervised with D. Okafor as the named approverThe manager, J. RidleyUp on evidence, down in one step
FeedbackEvery approve, edit and reject is recorded and usedAlways onFixed

The nine elements as they stand on the PO cancellation agent after the Friday release. The last column says when a change takes effect. Showcase estate.

Objective · the plannerThe objective is what the outcome check measures against. L. Tan wrote it as one sentence on the Friday, and every decision record points back to it. A new objective is a new version, because it changes what every other setting is for.
Data scope · the data leadThe scope names what the agent may read: purchase order lines, schedule lines, demand, stock and supplier class at plants 1210 and 1220. A. Novak owns it, because it is a question of which registered tables carry that meaning. Widening it is a new version.
Rules · the planner with the Planner AgentThe rules are the policy sentence and its planning table. L. Tan states the intent, the Planner Agent drafts the rows, and she adopts and edits them. A change is simulated on past data before it becomes a version.
Trigger · the plannerThe trigger says when the agent runs: 06:00 each working day, after the planning run, so it reads the night’s exceptions. It changes at once, because a different time moves the work and leaves the decision itself alone.
Actions permitted · the ownerThe actions are the short list of things the agent may do in your systems: draft a cancellation, send a change, open a query to the buyer. Everything outside the list stays outside it. A new action is a new version, with its own test write.
Limits · the owner and financeLimits are checked before every action. The value threshold of $25,000 belongs to L. Tan and J. Harrison in finance together; the daily budget of 60 and the confidence floor of 0.80 belong to the owner. Each limit carries its owner’s name and its version, and the audit record shows every check.
Escalation · the ownerEscalation says where a case goes when a limit hands it to a person: above the threshold, to D. Okafor as buyer, with the reason; a failed write-back, to the owner, L. Tan; a limit breach, the daily budget or the confidence floor, pauses the run and tells the owner. It changes at once, so a holiday or a new desk is a one-minute change.
Level · the managerThe level is the one setting a manager signs. J. Ridley set A2 with D. Okafor as the named approver on the Friday. It goes up on evidence: recorded cases pass the test, the owner approves by name, the escalation route is live. It comes down in one step, at once.
Feedback · always onFeedback is the one element you leave alone: it is always on. Every approve, edit and reject is recorded with the facts of that moment and read by the next cycle. The next section shows what one correction records.

Three owners, three kinds of change. The planner and the data lead change what the agent knows and looks for; the owner and finance change what it may do and how far; the manager changes how much of the work it carries. Each is a version, simulated or tested first. Only the trigger and the escalation route change at once, because they move the work and leave the decision alone.

Every action is a signal

Every approve, edit and reject teaches the next cycle

A correction is recorded with the facts of the moment, and the next version of the rule is built from the corrections that repeat.

The decision recordone per decision · kept01The agent draftsversion 3 on themorning’s facts02The personsignalsapprove · edit · reject ·override03The signal isrecordeddraft, decision, reason,facts, version04The outcome ischeckedagainst the objective, atday 1405Patterns arefoundcorrections that repeatbecome a proposal06A version isreleaseda named owner approves;the next run reads itthe arrows show one cycle · showcase estate
Six stations around one record. Two belong to a person: the signal at station 2 and the release at station 6. The next run reads the version station 6 released. Showcase estate.

A correction records both values: what the agent drafted and what the person sent, with the reason the person chose from a list and the facts as they stood at 06:00. A reason chosen from a list teaches faster than one typed by hand, because the same reason can be counted across decisions.

This is where the fourth cause from chapter 1 is answered. Corrections that live in an inbox are lost to the next cycle; corrections on the record are read by it. A buyer approving 40 drafts a day leaves 40 signals a day, and sees better drafts each week.

Edit · order 4500067990 · line 00030 · Monday 09:03
DraftedCancel 300 of 5513-B at plant 1220 · $7,160
SentCancel 200 · 100 kept for a confirmed order
Reason, from the listDemand confirmed after the run
Who and whenD. Okafor · 09:03 · role and session
Facts as read at 06:00excess 8.6 weeks · window open to Thu 24 Sep · standard
Rule versionPO cancellation policy · version 3

One of the 9 edits in six weeks, as the record holds it. Showcase estate.

Over six weeks the agent drafted 212 cancellations. D. Okafor approved 203 as written and edited 9, which is 96% approved as written, and those 9 edits are the signals the next version reads. Rule discovery also reads what buyers did before the agent existed: they had kept every line for one allocation-bound supplier, so Juno proposed an exclusion and j.park approved it as version 4. MIT NANDA, July 2025: 95% of pilots show a profit effect too small to measure, and the tools that retain feedback do better.

Two people, two settings

Finance and operations own different limits on the same agent

A limit carries the name of the team that set it, and the agent checks all of them before every action.

LimitSet byWhat it does on this agentOn the Monday
Value threshold · $25,000Finance · J. Harrison, with the ownerA line of $25,000 and above goes to D. Okafor as buyer, with the reason; at A2 a named person approves every line either way9 of 38 lines
Daily action budget · 60Operations · the ownerCaps the changes the agent may send in one day; at the cap the run pauses and the owner is told29 of 60 used
Confidence floor · 0.80Operations · the ownerA finding under the floor pauses the run on that line and reaches the owner with both readings; the recommendation stays emptyevery draft sat above it
Two-person rule · bank detailsFinance · J. HarrisonA change to bank details or payment terms waits for two named people, on any agent whose actions can reach those fieldsthis agent’s actions leave those fields untouched

Four limits on one agent, and the team that owns each. Showcase estate.

The threshold decides who sees a line first, and at A2 a named person approves every line either way. Finance co-owns it with L. Tan because it is a statement about money at risk. J. Harrison moves it with her as a new version, simulated first, while the rules and the trigger stay exactly as L. Tan set them.

The two-person rule runs the other way. Finance sets it once and it applies to every agent whose permitted actions could reach a supplier’s bank details or payment terms. The cancellation agent’s three actions leave those fields untouched, so the rule waits, and it would apply the day such an action were added.

Terms on this page
PatternOne of four ways a person and an agent share a piece of work, named by what passes between them.
SettingOne of the nine configurable elements of an agent, each with a named owner and a version.
SignalWhat a person does with an agent’s work: approve, edit, reject or override, recorded and read by the next cycle.
ExceptionA case a limit or a thin finding handed to a person, with the reason it was handed over.
Named approverThe person whose name goes on every approval at A2; D. Okafor on the PO cancellation agent.
Questions
Who changes a setting when two teams disagree?

Each element has one owner, named on the panel: finance for the value limits, the manager for the level, the owner for actions and escalation. A change is a new version with a name and a reason, on the record.

Does approving 29 drafts slow the buyer down?

The 29 took 22 minutes on the Monday, because the reason and the facts sat on each row. The work they replaced was the day of assembling that evidence by hand, which chapter 1 counted inside the 2.5 days.

What happens when the approver is away?

A named colleague receives the queue as cover, each item keeps its clock, and when a clock runs out the escalation route sends the item to the next desk. Both names go on the record.

Can a person override an agent at A3 or A4?

Yes. An override is a signal like an edit, recorded with its reason. A run of overrides is also the evidence that brings a level down one step, at once, by the owner or the manager.

Next chapter
The question this leaves

Settings are only worth having if turning them into a working agent is fast. How long does it take to make one?

Answered next

The PO cancellation agent, end to end

From one sentence in plain language to a released agent on a Friday, and what happened on the Monday after.

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.