Supply Chain and Operations
As of August 2026. Phase 6. The department the founding metaphor came from.
1. The scenario
A distribution business runs an ERP for planning, a WMS in the warehouse, transport management, a supplier base reached by EDI and portals, and a planning team that spends most of its week reconciling exceptions: short shipments, late suppliers, allocation conflicts, and demand signals that disagree.
2. Agent team design
- Agent
- Exception resolution agent
- What it does
- Investigates supply and order exceptions, gathers evidence across systems, proposes disposition
- A x L position
- A3, L2
- Notes
- The core of the department's value
- Agent
- Supplier communication agent
- What it does
- Chases, confirms and reconciles with suppliers across email and portals
- A x L position
- A3, L1
- Notes
- External-facing. Disclosure applies
- Agent
- Planning analysis agent
- What it does
- Explains plan variance and demand-signal disagreement with provenance
- A x L position
- A2, L2
- Notes
- Explains; the planner decides
- Agent
- Document processing agent
- What it does
- Extracts and validates shipping, customs and compliance documents
- A x L position
- A3, L1
- Notes
- Parsing fidelity is the ceiling here
3. Planes activated
Action (direct: ERP, WMS and EDI through wrapped APIs), Knowledge (direct: parsing fidelity caps everything), Improvement (direct), Human, Control, Evidence, Execution (supporting).
4. Controls
- Idempotency and compensation on every agent-initiated action. A duplicated purchase order is a real cost, and retry behaviour is where duplication comes from.
- Approval gates by consequence class: quantity and date changes below a threshold flow, commercial terms do not.
- Supplier-facing communication carries disclosure and is retained.
- Parsing provenance at page and cell level, because a customs document error is a compliance event.
- Degraded-mode behaviour specified: the warehouse does not stop when an agent does.
5. Economics
Per run. Moderate, with tool fan-out across systems.
Per resolved outcome. Cost per exception cleared, against the planner's loaded hour. The compounding value is upstream: every exception whose root cause is recorded improves the master data everything else grounds on, which is the slow loop that matters more than the fast one.
6. Honest limits
- No named-factory case exists of an MES agent taking autonomous production action with measured outcomes. Widely circulated figures in this space trace to content farms. This blueprint stops at the planning and administrative layer for that reason.
- Warehouse and shop-floor actuation is R05 territory and carries the OT boundary: agents on the information path, not the control path.
- Parsing quality caps document processing. Even the best parsers lose at least 14% of retrieval performance against ground-truth structure, so document agents need spot-check metrics rather than trust.
- Master data quality is usually the real constraint, and no agent fixes it faster than it degrades without an owner.
7. Metrics
Exception clear rate and rework rate. Duplicate-action rate, which should be zero. Parsing fidelity spot checks. Master data disagreement count, trending down. Supplier response cycle time. Planner hours recovered.
Sources
research/R03-integration-fabric/ (idempotency, compensation, EDI), research/R04-systems-of-record/, research/R05-lob-and-ot/ (the OT boundary and the MES finding), research/R14-agent-data-engineering/ (parsing).
Source: blueprints/departments/supply-chain.md in the evidence repository behind this site.