Agentic AI Chapter 7 of 9

A language model helps write the plan; the rules engine runs it.

The last chapter left one question open: where the language model sits, and what it gets to see.

FRIDAY · DESIGN TIME · THE CLOUD MODELtoken cost applies here09:00 · L. Tan states the intentone sentence, in plain language, to the Planner Agent09:40 · the cloud model drafts five rowsthree checks, one action, one limit, from a fixed vocabulary10:10 · she adopts four rows, edits onethe excess check, in her own words: version 211:00 · limits, level and approver$25,000 · 60 a day · floor 0.80; level A2; approver D. Okafor11:30 · simulation on last quarter184 lines · $1.9M of open value · 31 routed to the buyer13:30 · J. Ridley publishes version 3planning tables in the versioned rule storeMONDAY · RUN TIME · INSIDE YOUR ESTATEzero token cost06:00 · the rules engine reads version 3the same rule gives the same result on every runThe tools read the facts in place38 lines from SAP S/4HANA and Oracle FusionThe model in your estate words the basiszero token cost; the rows stay inside09:12 · D. Okafor has approved the queue29 drafts worth $187,300, each by name, in 22 minutes09:12 to 09:21 · write-backSAP acknowledges each change; every answer stored word forwordOne record per decisionDEC-2026-09-21-0417 · eight partsversion 3: theplanning tables
The Friday on the left, the Monday on the right, and version 3 as the handoff between them. Token cost applies on the left alone. Showcase estate.
2
halves: the Friday writes the plan with a cloud model; the Monday runs it inside your estate
showcase estate
zero
token cost across the cancellation agent’s 40 runs in eight weeks
showcase estate
8
parts in the record of every action, from signal to outcome
product
100%
of actions traced to a rule in the production PO cancellation agent
production
So far
  1. How it worksYou have seen the six stages and the ontology at their centre.
  2. The PO agentYou have seen the Planner Agent draft five rows from one sentence on the Friday, and the rules engine run version 3 on the Monday.
  3. People and agentsYou have seen the limits and the audit record among the settings.
Design time and run time

The cloud model works on the Friday; the rules engine works on the Monday

Two halves, one handoff: the planning tables released as version 3.

Go back to the Friday of chapter 5. At 09:00 L. Tan typed one sentence: “Cancel open purchase order lines where projected excess is above eight weeks of demand, the line is inside its cancellation window, the line value is under $25,000 and the supplier class is standard. For lines of $25,000 and above, draft the cancellation and ask the buyer.” A cloud language model of the company’s choice read it, and five rows came back.

That is design time, the one place a cloud model appears: it reads the intent and drafts rows from a fixed action vocabulary. Token cost applies here: $420 over 62 calls to author the policy, $260 over 38 calls to simulate and refine it.

Now the Monday. At 06:00 the Lima Rules Engine read version 3 from the rule store and evaluated it on the 38 lines. For the basis under each draft it called the small language model inside your estate. The agent’s 40 runs over eight weeks carried zero token cost.

The language model drafts, explains, discovers candidate rules in past planner behaviour and detects drift. The rules engine decides and executes, so every write-back traces to a versioned rule and, at A2, to D. Okafor.

Token manager · PO cancellation agent · eight weeks from 18 September 2026
Friday · authoring the policy · 62 calls$420
Friday · simulating and refining it · 38 calls$260
Six weeks · rule discovery · 112 calls$604
Eight weeks · 40 runs of the agentzero token cost

Showcase estate. Bars are shares of the agent’s $2,500 budget; $1,284 over eight weeks, all at design time.

The boundary

Names and words may cross at design time; the rows stay inside

The line is drawn on the model link and checked on every call.

INSIDE YOUR ESTATEEXTERNAL MODELWHAT STAYSOrder rows and schedule linesPrices and contract termsCustomer and supplier namesStock and demand figuresWHAT MAY CROSSTable and column namesYour vocabulary and acronymsThe sentence the planner typedThe action vocabularyBOUNDARYThe model you choseGemini, Claude, OpenAI,Llama or anotherDesign timethe Friday, while the planis writtenMeteredbudgeted and counted percalla model link serves after a live handshake confirms the model, the key and this boundary
The middle column is everything that crossed on the Friday. The 38 lines and their values stayed on the left on the Monday. Showcase estate; the rule is the product’s.

On the Friday the cloud model needed to know what the tables are called and which words the company uses. So the names of the registered tables and columns crossed, with the vocabulary P. Sharma keeps (NCNR, the lead time fence, “strat supplier”) and the sentence L. Tan typed.

On the Monday the 38 lines, their values, the stock and demand figures behind the excess and the suppliers’ names stayed in SAP S/4HANA and Oracle Fusion, read in place. They stayed inside your estate when a basis was worded.

S. Iyer chooses the modelGemini, Claude, OpenAI, Llama or another; the key stays in your key store.
Limits

Five limits are checked before an action leaves the agent

Each check is recorded, and a limit that holds an action turns it into a draft for a person.

LimitWhat it doesOn the cancellation agentOn the Monday
Value thresholdLines under the value become drafts; lines above it go to a person.$25,000 a line29 lines under it became drafts worth $187,300; 9 above it reached D. Okafor as buyer, for his judgement, with the reason for each.
Daily action budgetCaps the actions one agent may take in a day.60 actions a dayThe run used 29 of 60; the counter sits in every record.
Confidence floorHolds a draft under the floor; a person receives the finding.0.80Line 00020 of order 4500067123 scored 0.91 and passed.
Rate limitSpaces the writes an agent sends to a source system.12 writes a minuteThe 29 approved writes reached SAP S/4HANA over 9 minutes, each acknowledged.
Two-person approvalCalls for a second named approver on sensitive fields.Bank details · payment termsA cancellation changes a quantity and a deletion indicator, so this rule stayed idle.

You met the limits in chapter 4 as settings the owner and finance own. The rules engine reads all five before every action and writes each check into the record. They also made the level review safe: when J. Ridley moved lines under $25,000 to A3 after six weeks, the threshold, the budget and the floor stayed where they were, so the reach of that change stayed small.

The audit record

One decision, read the way an auditor reads it

DEC-2026-09-21-0417 holds eight parts; each answers one question a reviewer asks.

PartWhat the record holdsWhat a reviewer checks it against
1 · SignalException message 0034921 for material 7741-X at plant 1210, read Monday 06:00The message in the planning suite
2 · ReasoningVersion 3, released 18 September 2026 by J. Ridley; the facts as read at 06:01: projected excess 11.4 weeks · line value $8,420 · window closes Thursday 24 September · supplier class standardThe rule text of version 3; the values the sources held at 06:01
3 · RecommendationDrafted 06:06: cancel line 00020 of order 4500067123; three conditions matched; confidence 0.91The three conditions against those facts
4 · Limits checkedValue under $25,000 passed · daily budget 29 of 60 · confidence floor 0.80 passedThe limits set in version 3
5 · ApprovalD. Okafor opened the queue at 08:50 and approved it as drafted at 09:12, after 3 h 06 min in the queue; session and role recordedThe named approver; the level in force, A2
6 · ActionPurchase order change posted through the released interface for this agentThe write route opened after the test write
7 · AcknowledgementOrder 4500067123 item 00020 carries the deletion indicator; message 20260921-114 stored as returnedThe document in SAP S/4HANA itself
8 · OutcomeProjected cover 6.2 weeks at day 14; the objective was metThe objective on the agent; the cover on day 14

Showcase estate. A rejected draft and a run at A4 leave the same eight parts.

Chapter 5 showed this record as the run wrote it. An auditor reads it the other way, from the outcome back to the signal. Each part is checked against something outside Juno: the document in the ERP, the rule text, the values the sources held that morning.

The seventh part carries the most weight, because it is the ERP’s own word: the document number and the message SAP returned, stored as received. A reviewer opens the order and finds the same item.

Point-in-time replayEvery governed object carries an append-only version history, so the rule, the mapping and the vocabulary are rebuilt as they stood that day.
Retention you setRecords stay on your own storage for the period your policy sets; the EU AI Act, Article 26, asks deployers for at least six months.
The layers

Four layers, named part by part, on the cloud you already run

Open-source parts, named layer by layer, so an architecture review has something to check.

THE PLATFORM IN FOUR LAYERSPEOPLEChatLimaquestionsApproval queuesSmartAppsManager reportsper 1,000 decisionsAPIs and BIone definitionAGENTSPlanner Agentintent to a planRules enginedecides, executesRule storeversionsToken managerbudgets, routingMEANINGOntologyobjects, measuresKnowledge indextables, valuesVocabularyyour termsMappingscolumn to attributeDATAYour sourcesread in placeFederated queryacross sourcesGraph databasethe ontologyObject storagefiles, archivesWhere people meet itAsk, approve, review. Each screen shows theterms it used and the rule that produced theaction.Where decisions are madeThe rules engine decides and executes. Themodel drafts, explains, discovers rules anddetects drift.Where words get meaningOne definition per object and per measure, readby the conversation and by every agent.Where your rows staySources are read in place; the rights of theconnection and of the asking person apply toevery read.
Each layer has one job and can be changed on its own; the meaning layer keeps a question, a rule and a report on one definition. Product.

Read downwards, the picture answers the security committee. A person asks or approves at the top; the agents layer decides and executes; the meaning layer supplies the one definition of a purchase order line that chapter 6 put at the centre; the rows stay in your systems until a query reads them. Read upwards, it answers the architect: a new source joins at the bottom and reaches every screen through one mapping and one set of access rules.

LayerWhat it is built fromWhat that gives you
InterfaceReact, Node.js and NginxOne interface for registering, asking, reviewing and approving.
ModelsGemini, Claude, OpenAI, Llama or another, interchangeableYour choice of cloud model at design time; a small model inside your estate at run time.
ServicesSpring Boot and FastAPI microservices, in containersEach service scales on its own; an update leaves the rest running.
Storage and queryMariaDB, an ArangoDB graph database, a federated query engine, RabbitMQ and object storageThe ontology as a graph, the registry and the queue; your sources read in place.
Run timeKubernetes, Docker and LinuxOne deployment model across the four cloud targets.

Google Cloud

Containers on managed Kubernetes, with the cloud’s own storage and identity.

Microsoft Azure

The same containers and deployment model, under your tenancy.

Amazon Web Services

The same again, inside your account’s network and key management.

A private cloud

Your own data centre on Linux, for estates that keep everything inside.

Security and frameworks

Twelve layers of security, and the frameworks an analyst will check against

Identity, role-based access and the asking person’s rights apply to every answer and every action.

Twelve security layersFrom the session to the infrastructure: transport encryption, encryption at rest, network boundaries, secrets in your key store, audit logging.
Single sign-on and role-based accessPeople arrive through your identity provider; rows and columns are granted by role, and the grant applies to every query and every agent run.
The asking person’s rights on every answerChatLima answers as the person who asked and names the scope their role allows, as L. Tan’s Wednesday answer did with plants 1210 and 1220.
What buyers and auditors ask forFramework, with its dateThe Juno control
Human oversight. People can oversee the system, interpret its output and override it.EU AI Act, Article 14 on human oversight; the Act came into force on 1 August 2024.The level, set per agent: a named approver at A2, exceptions at A3, audit at A4.
Logs. Events are recorded automatically across the life of the system.EU AI Act, Article 12 on record-keeping.Eight parts for every run, written as it happens, kept for the period you set.
The ability to stop. A person can halt the system at any moment.EU AI Act, Article 14, which asks that overseers can stop a system.One step down, at once, by the owner or a manager; the daily action budget holds the reach of a run.
Accountability. Trained overseers are named, the system is watched, logs are kept for at least six months.EU AI Act, Article 26 on the duties of deployers.A named owner on every agent, a level register with the evidence behind each change, and manager reports per 1,000 decisions.
Risk management. Risk is identified, measured and managed while the system runs, inside a system a third party can audit.NIST AI Risk Management Framework 1.0, 26 January 2023: govern, map, measure, manage. ISO/IEC 42001, published 18 December 2023.Simulation on past data, accuracy testing, outcome checks, drift detection, and a version history on every governed object.
Where this sits today

Most supply chain planning agents sit outside the Annex III list of high-risk uses in the EU AI Act. Regulation (EU) 2026/1744, published 24 July 2026, moved the stand-alone high-risk duties to 2 December 2027. Buyers and auditors ask for the Article 12, 14 and 26 pattern all the same, and Juno follows it for every agent. Your own counsel makes the legal reading; dates checked 19 September 2026.

Terms on this page
Design timeThe half where a person states intent and a cloud model helps draft, simulate and publish a plan.
Run timeThe half where the rules engine executes a published plan inside your estate, at zero token cost.
Model linkThe governed record of a provider, a model, a key reference and the boundary rule.
Model boundaryThe rule on a model link: names and words may cross at design time; row data stays inside.
Decision referenceThe identifier that holds the eight parts of one action, from signal to outcome.
Questions
What reaches the cloud model, exactly?

At design time: the names of your tables and columns, your vocabulary and the sentence the planner typed. Order rows, prices, stock figures and names stay inside your estate.

Can we stop an agent once it is running?

Yes. An owner or a manager lowers its level in one step and the change applies at once. The agent carries on at the lower level, drafting for a person.

Which model does Juno use, and can we change it?

The one you choose: Gemini, Claude, OpenAI, Llama or another. A swap leaves the planning tables, the ontology and the audit trail as they are.

The question this leaves

You have seen one agent built and run. What else can be built this way, and where should you start?

Answered next

The use cases

Five end-to-end chains, the use cases the market is deploying now, and the map of everything else.

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.