On Friday 18 September 2026 L. Tan typed one sentence into Lima Juno, and by 13:30 the PO cancellation agent was released at A2 with D.
The Friday of the first week, hour by hour: sentence, question, five rows, one edit, limits, level, simulation, release.
At 09:00 L. Tan opened the Planner Agent and typed the policy in the words the desk uses: purchase order line, projected excess, cancellation window, supplier class, each a term the estate mapped on the Tuesday or taught on the Thursday.
The Planner Agent asked one question, because the sentence left one choice open: which demand the excess is measured against. She answered in a line. At 09:40 five rows came back, three checks, one action and one limit, each naming the data it reads.
The rows come from a fixed action vocabulary, the actions Juno supports for those objects, so the plan stays inside what the rules engine can carry out and the audit trail can record.
One question before I draft. Projected excess compares supply with demand. Should demand be the forecast alone, or the forecast plus open sales orders?
Five rows drafted: three checks, one action and one limit, each with the data it reads. Adopt them one by one or together, edit any row, then simulate.
| Row | Kind | What it tests or does | Data it reads | Then |
|---|---|---|---|---|
| 01 | Check | Projected excess is above 8 weeks of demand edited by L. Tan · 10:10 | stock on hand, stock on order, forecast, sales orders | carry on |
| 02 | Check | The cancellation window is open on the run date | order date, cancellation window in days | carry on |
| 03 | Check | The supplier class is standard | supplier record, supplier class | carry on |
| 04 | Action | Draft the cancellation of the open quantity | open quantity, line value, currency | row 05 decides the route |
| 05 | Limit | Line value against $25,000 | line value | under: the queue · at or above: the buyer |
She adopted four rows and edited one. Row 01 measured excess from stock on hand alone; she added stock on order, so the projection counts supply already on its way. The draft is version 1, her edit version 2, and the limits and level added at 11:00 make version 3.
Every setting has a value, an owner and a rule for how it changes; the replay shows what those values would have done.
| Element | Setting | Set by | Changes |
|---|---|---|---|
| Objective | Cancel purchase order lines that stand above the need, inside their cancellation window | L. Tan, planner | New version |
| Data scope | Purchase order lines, schedule lines, demand, stock and supplier class at plants 1210 and 1220 | A. Novak, data lead | New version |
| Rules | The policy sentence and its planning table, version 3 | L. Tan with the Planner Agent | New version, simulated first |
| Trigger | 06:00 each working day, after the planning run | L. Tan, planner | At once |
| Actions permitted | Draft a cancellation · send a change · open a query to the buyer | L. Tan, owner | New version |
| Limits | Value threshold $25,000 · daily action budget 60 · confidence floor 0.80 | L. Tan and J. Harrison, finance | New version |
| Escalation | Above 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 owner | L. Tan, owner | At once |
| Level | A2 Supervised with D. Okafor as the named approver | J. Ridley, manager | Up on evidence, down in one step |
| Feedback | Every approve, edit and reject is recorded and used in the next cycle | Always on | Fixed |
The settings screen at 11:00: each row names its owner and how a change takes effect. Showcase estate.
Chapter 4 listed the nine settings and who owns each; here they carry values, and the owners are people the reader has met. L. Tan owns the agent: the objective, the trigger, the actions and the escalation route; A. Novak owns the data scope; J. Ridley sets the level and releases each version, and J. Harrison from finance the $25,000 threshold.
The level is J. Ridley’s as the manager: A2 supervised, with D. Okafor as the named approver. Chapter 3 gave the definition: the agent decides and drafts, and a named person approves each action.
| What the replay found | Lines | What it means |
|---|---|---|
| Matched all three checks | 184 | $1.9M of open value the policy would have released |
| Under $25,000 | 153 | would have reached the approval queue as drafts |
| $25,000 and above | 31 | would have reached the buyer with the reason for each |
| Kept by the policy, cancelled by hand at the time | 2 | both outside their window; J. Ridley read both before he signed |
The replay ran from 11:30 and reads like a test the desk can mark. Showcase estate.
Simulation lets a manager sign. Last quarter was replayed beside what people did at the time: 184 lines matched, worth $1.9M, and 31 would have gone to the buyer. The two lines the policy would have kept were outside their window when the desk cancelled them by phone. J. Ridley read both and published version 3 at 13:30.
The run, the split, the queue and the confirmation, in the order they happened.
This is the Monday chapter 1 opened on, seen from the agent’s side. At 06:00, after the planning run, the rules engine read version 3 and gathered its facts in place from three systems. Of 212 open lines, 38 passed the three checks, worth $412,000.
Row 05 then did what the sentence says. Twenty-nine lines under $25,000 became drafts in D. Okafor’s approval queue, worth $187,300. Nine at or above the threshold went to him in his buyer’s chair, with the reason and the value, for his judgement before the window closed on Thursday 24 September.
A named person decided every one of the 38; the threshold decided only how each line reached him. That is A2 in practice, and it is why the six-week record below could carry the level up.
The four steps chapter 1 promised add up on this morning: the question answered in 4.2 seconds, 29 cancellations drafted in 6 minutes, the queue approved in 22, the write-back confirmed in 9. Thirty-eight minutes, against the old path’s 2.5 days.
| Line | What the agent proposes | Value | Basis | Confidence | Decision |
|---|---|---|---|---|---|
| 4500067123 · 00020 | Cancel 200 of 7741-X at plant 1210 | $8,420 | excess 11.4 wk · window Thu 24 Sep · standard | 91% | Approved 09:12 · DEC-0417 |
| 4500068004 · 00020 | Cancel 1,200 of 8802-K at plant 1210 | $24,100 | excess 9.3 wk · window Wed 23 Sep · standard | 93% | Approved 08:58 |
| 4500067990 · 00030 | Cancel 300 of 5513-B at plant 1220 | $7,160 | excess 8.6 wk · window Thu 24 Sep · standard | 88% | Edited to 200 · approved 09:03 |
| 4500068110 · 00010 | Cancel 150 of 9021-M at plant 1220 | $5,900 | excess 12.1 wk · window Fri 25 Sep · standard | 90% | Approved 09:05 |
| 4500068221 · 00050 | Cancel 80 of 4410-T at plant 1210 | $2,280 | excess 10.2 wk · window Thu 24 Sep · standard | 90% | Approved 09:10 |
| 24 more lines | each under $25,000 | $139,440 | three conditions matched on each | All approved by 09:12 |
D. Okafor’s queue at 08:50: the basis in the words of the policy, the confidence and the value at stake on every row. Showcase estate.
Twenty-nine confirmed changes, and for each of them a record that starts at the exception message and ends at day 14.
Approval sent the 29 changes to the ERP through the purchase order interface it publishes, under the agent’s own account. The ERP ran its own checks and answered each with a document number; Juno read every line back before closing the record. Nine minutes, confirmed.
Every decision leaves one record on the audit trail. The ledger shows one of the 29: line 00020 of order 4500067123, the controller board 7741-X at plant 1210, from the exception message at 06:00 to the outcome check at day 14.
Its eight parts are the ones chapter 2 named in the anatomy and chapter 4 listed among the settings, from the signal to the outcome, with the rule version and the facts as read inside the reasoning.
Each was written as it happened, in an append-only store, so an auditor can read the chain from the change in the ERP back to the sentence L. Tan typed on the Friday.
| When | Part of the record | What decision DEC-2026-09-21-0417 holds |
|---|---|---|
| Mon 06:00 | Signal | Exception message 0034921: material 7741-X at plant 1210 stands above the need after the customer push-out. Line 00020 of order 4500067123, read from SAP S/4HANA. |
| Mon 06:01 | Reasoning | Version 3, released 18 September 2026 by J. Ridley. Facts as read: projected excess 11.4 weeks · window closes Thu 24 Sep · supplier class standard · line value $8,420. Rows 01 to 03 matched. |
| Mon 06:06 | Recommendation | Cancel 200 of 7741-X, the open quantity on the line. Confidence 0.91. Row 05: under $25,000, so the draft goes to the approval queue. |
| Mon 06:06 | Limits checked | Value threshold $25,000: passed · confidence floor 0.80: passed · daily action budget: 29 of 60 after this run. |
| Mon 09:12 | Approval | D. Okafor approved as drafted after 3 h 06 min in the queue; name, role, session and minute recorded. |
| Mon 09:14 | Action | Purchase order change sent to SAP S/4HANA through the released purchase order interface, under the agent’s own account, with the document version it had read. |
| Mon 09:14 | Acknowledgement | Order 4500067123 item 00020 carries the deletion indicator; message 20260921-114 stored as returned; the read-back shows an open quantity of 0. |
| Mon 5 Oct | Outcome | Day 14: projected cover on 7741-X at plant 1210 read again, 6.2 weeks against 11.4 on the Monday; verdict met against the objective. |
Showcase estate. One decision reference holds all eight parts; chapter 7 reads this same record the way an auditor does.
A test write, a change set and a release on the Friday; scans, re-runs, reviews and reports from then on.
Deploying meant three things. First, a confirmed test write. At 12:30 on the Friday the agent changed one purchase order line in the test system of each ERP, from its own account; S. Iyer and J. Ridley read the answer, and write-back was switched on for purchasing. That test is the switch, per application and per object.
Second, two environments and a change set. The draft, the edit, the limits, the level and the simulation happened in the build environment while the live agents kept working. At 13:30 J. Ridley released them as one change set, and the release log kept his name and the minute.
Third, the release itself. The first run that Friday afternoon drafted six lines for D. Okafor’s queue, so the desk saw the agent at work before the weekend; the Monday run at 06:00 was its second.
From then on four things keep running on their own, each a setting or a report a person reads: the quiet part of deployment, and what keeps a released agent right as the estate changes.
The approvals moved the level, in the four steps the ladder sets, and the threshold stayed.
Over six weeks and 30 runs the agent drafted 212 cancellations. D. Okafor approved 203 as written and edited 9, which is 96% as drafted. The line J. Ridley had set for a move to A3 was 95% approved as written, held for four consecutive weeks; weeks 3 to 6 held it. Chapter 4 explained why the drafts improved: every edit was recorded and used, so the proposed quantities moved towards the ones the buyer chose.
That record is what the ladder asks for. Chapter 3 gave the four steps of a level change: recorded cases pass the test, the owner approves by name, the escalation route is live, and the change goes on the audit trail. On Friday 30 October J. Ridley took them in that order.
He moved lines under $25,000 to A3 Delegated: the agent now executes those cancellations inside version 3, and D. Okafor handles exceptions only. Lines at or above the threshold still arrive as drafts for the buyer, because the value threshold stayed where finance put it.
The blueprint designs this role at A3; the showcase started it at A2. Six weeks of evidence carried it up one rung, and one step would bring it down again. That is the ladder working as designed.
| Setting | Before the review | After the review | What carried it |
|---|---|---|---|
| Level, lines under $25,000 | A2 supervised · D. Okafor approves each draft | A3 delegated · the agent executes inside version 3; D. Okafor handles exceptions only | 203 of 212 approved as written; the 95% line held four consecutive weeks |
| Level, lines at or above $25,000 | A2 · drafted for the buyer with the reason | A2 · drafted for the buyer with the reason | the value threshold stays |
| Escalation route | above 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 owner | the same three routes, unchanged | live before the level moved |
| Approved by name | J. Ridley · Fri 30 Oct 2026 · 16:10 | on the audit trail as a level change | 212 recorded cases, the 9 edits read one by one |
The level review as the audit trail records it. Showcase estate.
Four production facts, and a showcase week that shows how they come about.
The cancellation policy is written in plain language by the planners who own it.
productionThe policy runs on purchase order data read across the company’s ERP systems, in place.
productionWrite-back happens under human approval.
productionEvery action traces to a rule, and run time carries zero token cost.
productionThe agent this chapter followed is the one role from the blueprint that runs in production today, at a high-tech manufacturer that keeps its name private. Four facts carry the production label, and this page carries them under that agreement.
Planners author the cancellation policy in plain language, and the rules engine executes it on purchase order data read across the company’s ERP systems. Write-back happens under human approval. And 100% of actions trace to a rule, at zero token cost at run time.
The Friday, the Monday and the six weeks belong to the showcase estate, and their figures are illustrative. They are there so a reader can follow every step: how such an agent is configured, how it works and how it is deployed.
The mechanism is the same in both places: a sentence, a table, a simulation, a version, a named approver, a confirmed write-back and one record per decision.
Four and a half hours, from the sentence at 09:00 to the release at 13:30. Authoring cost $420 over 62 model calls and the simulation $260 over 38; the 30 runs of the six weeks cost zero tokens.
The message is stored in the ERP’s own words, the action waits, and the approver sees the reason on the row. The next run raises the same line while its window is still open.
Each setting has an owner, as chapter 4 set out. The trigger and the escalation route change at once; the objective, the scope, the rules and the limits become a new version, simulated first; the level moves on evidence.
This agent read purchase order lines from SAP and from Oracle and treated them as one thing. How did it know?
Six stages, one shared map of your supply chain, and why the same machinery serves questions and agents alike.
Each one builds on the one before it.
Reads your rows in place.
Connect and move ↗One record per real thing.
Master and govern ↗Apps for every decision.
Operate and decide ↗Asks and acts, at your level.
Ask and act ↗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.
