Agentic Architecture · Supply Chain

Supply Chain Agentic Demo

System architecture for a one-hour client demo — seven planning agents across a planning platform and a lakehouse.

Agentic Use Cases

Live Click-Through Demos

Narrated Video Walkthroughs

~7d

Build Timeline

Platforms

System Architecture

flowchart TD
  subgraph DATA["Data Sources"]
    KIN["Kinaxis Maestro\nlive instance"]:::source
    SYNTH["Synthetic Data\nWIP files · yield curves\npartner portals"]:::source
  end
  subgraph BRIDGE["MCP Bridge"]
    MCP["MCP Client\nKinaxis Agent Studio"]:::source
  end
  subgraph ENGINE["Claude API Reasoning Engine"]
    REASON["Agent Reasoning Layer\nprompt engineering · tool definitions\nstructured outputs"]:::source
  end
  subgraph AGENTS["Agent Layer — 7 Use Cases"]
    A6["UC6 · Supply Disruption\nLIVE"]:::live
    A8["UC8 · Inventory Expiry\nLIVE"]:::live
    A1["UC1 · WIP Recon"]:::video
    A2["UC2 · Predictive Yield"]:::video
    A5["UC5 · Parameter Opt"]:::video
    A7["UC7 · Allocation"]:::video
    A3["UC3 · NPI Phase-Gate"]:::stretch
  end
  subgraph UI["Demo UI — Split Screen"]
    ACT["Agent Actions\nKinaxis screens\nMaestro scenarios"]:::actions
    REAS["Agent Reasoning\nplain language\nevidence attached"]:::actions
  end
  subgraph HUMAN["Human-in-the-Loop"]
    APPROVE["Planner Accept / Reject\ncommit · override · escalate"]:::source
  end
  subgraph TARGET["Planning & ERP Systems"]
    MAESTRO["Kinaxis Maestro\nplanning data · scenarios"]:::source
    ORACLE["Oracle ERP\npurchase orders · items"]:::source
  end

  KIN --> MCP
  SYNTH --> MCP
  MCP --> REASON
  REASON --> AGENTS
  A6 --> ACT
  A8 --> ACT
  A1 --> REAS
  A2 --> REAS
  ACT --> APPROVE
  REAS --> APPROVE
  APPROVE -->|"committed"| MAESTRO
  APPROVE -->|"committed"| ORACLE

  classDef source fill:#FFFFFF,stroke:#CDBBA8,stroke-width:1.5px,color:#241A12
  classDef live fill:#E4F1EE,stroke:#0E7C6E,stroke-width:2px,color:#241A12
  classDef video fill:#FAF0DC,stroke:#A5650A,stroke-width:1.5px,color:#241A12
  classDef stretch fill:#F7F3EF,stroke:#E1D6CB,stroke-width:1px,color:#574A3C
  classDef actions fill:#E9E7FB,stroke:#4F46E5,stroke-width:1.5px,color:#241A12

live demo

video walkthrough

split-screen panel

Split-Screen Demo UI — Build Requirement

The split screen is a build requirement, not a presentation trick: agent actions on one side, reasoning on the other, paced for a human to follow.

Agent Actions

01Query Kinaxis for scheduled receipts, pegging, BOM data

02Detect disruption: commit date moved, inventory at risk

03Model 2–3 mitigations as Maestro what-if scenarios

04Present options with penalty arithmetic visible

05Execute chosen scenario on planner approval

Agent Reasoning

01Plain-language root cause: “Commit moved because cycle time at OSAT slipped 4 days”

02Named customers affected, with revenue + margin at stake per customer

03Lateness penalty arithmetic: 2 minor misses > 1 severe miss

04Evidence chain: source data → analysis → recommendation

05Before-state: planner + 50 workbooks → seconds

Agent Inventory — 7 Use Cases

Priority	#	Use Case	Platform	Type	Demo Message	Handoff
P1	`6`	Supply Disruption Exception Handler	Kinaxis	LIVE	Why the commit moved, in plain language, in seconds	← Receives from UC1
P2	`8`	Inventory Expiry & Alternate BOM	Kinaxis	LIVE	At-risk lots converted to revenue before write-off	—
P3	`1`	Agentic WIP Reconciliation	Kinaxis + Databricks	VIDEO	Receipt dates corrected before the miss, not after	→ Feeds UC6, UC7
P4	`2`	Predictive Yield Agent	Databricks	VIDEO	Demonstrated yield beats stated yield, with evidence	↔ Overlaps UC5
P5	`5`	Planning Parameter Optimization	Kinaxis	VIDEO	Parameters self-heal inside Maestro, not in a side tool	↔ Overlaps UC2
P6	`7`	Agentic Allocation	Kinaxis	VIDEO	Constrained supply reallocated with defensible rules	← Receives from UC1
P7	`3`	NPI Phase-Gate Agent	TBD	VIDEO	Agent layer over the build already in flight	—

Agent Handoff Map

Visible agent handoffs are the compounding argument: WIP reconciliation feeds exception handling and allocation; yield monitoring overlaps parameter audit.

flowchart LR
  subgraph LIVE["Live Demos — Full Click-Through"]
    UC6["UC6\nSupply Disruption\nException Handler"]:::live
    UC8["UC8\nInventory Expiry\nAlternate BOM"]:::live
  end
  subgraph VIDEO["Video Walkthroughs — ~90s Each"]
    UC1["UC1\nAgentic WIP\nReconciliation"]:::video
    UC2["UC2\nPredictive Yield\nAgent"]:::video
    UC5["UC5\nPlanning Parameter\nOptimization"]:::video
    UC7["UC7\nAgentic\nAllocation"]:::video
    UC3["UC3\nNPI Phase-Gate\nAgent"]:::stretch
  end

  UC1 -->|"WIP flush\ntriggers exception"| UC6
  UC1 -->|"shortfall from\nreal WIP data"| UC7
  UC2 -.->|"yield monitor overlaps\nparameter audit"| UC5

  classDef live fill:#E4F1EE,stroke:#0E7C6E,stroke-width:2px,color:#241A12
  classDef video fill:#FAF0DC,stroke:#A5650A,stroke-width:1.5px,color:#241A12
  classDef stretch fill:#F7F3EF,stroke:#E1D6CB,stroke-width:1px,color:#574A3C

Data Layer Architecture

Live Data

Kinaxis Maestro

Via an MCP client bridge to Kinaxis Agent Studio. Real-time queries against scheduled receipts, pegging, BOM, where-used, and Maestro what-if scenarios.

Synthetic Data

Domain-shaped Files

Synthetic is explicitly allowed by the brief, but must mirror part number formats, bin grades, and OSAT WIP file layouts so a planner never sees data that looks fake.

External Signals

Simulated Triggers

Pre-computed events for demo realism: supplier delay alerts, commit date changes, WIP receipt mismatches, and yield gap detection.

Lakehouse strategy: Live demos run entirely on the planning platform. Lakehouse-sourced data is synthetic. The yield walkthrough uses pre-computed yield curves with the reasoning layer on top — the concept lands without live infrastructure.

Architecture Decisions

Reasoning

Claude API

Prompt-engineered reasoning chains with structured outputs. Evidence-attached responses — the agent shows its work, not just its answer.

Kinaxis-first

Not a custom UI

Anchor every demo in Kinaxis or Databricks screens. A custom UI reads as a prototype; their own systems read as a deliverable.

Human-in-the-loop

Proposed vs committed

Every action that changes planning data requires planner approval. The agent explores paths; the human chooses the destination.

Slow motion

Paced for humans

Autonomous agents are invisible on screen. Actions appear incrementally, reasoning unfolds step by step, mitigations compared before committing.

Before-state

In numbers

Every demo closes with the contrast: planner + 50 workbooks → seconds, 5–6 clicks → single ranked list. Speed and legibility are the message.

Discovery first

Client-specific

A short discovery pass makes demos client-specific: part formats, bin grades, file layouts. Generic demos lose planners in 30 seconds.

Client Session Format — 60 Minutes

Live Demos · ~15 min each

Full Click-Through

Real interaction and real data flow. The planner drives the experience; the agent responds in real time. Enough detail that the client sees the thing is close to ready.

Video · ~90 sec each

Narrated Walkthroughs

Concept walkthroughs, recorded not driven live. Show the architecture, the agent’s capabilities, and the before/after contrast. Close with proof points.

The Objective

Show the System, Not One Tool

The architecture is the story: seven agents that compound through visible handoffs. A planner leaves understanding the whole system, not a single feature.