Chapter 04
How agents actually work
When a disruption hits, one Main Operations Agent fans out five narrow-scoped sub-agents at once and returns the decision with its full reasoning on display.
Incoming disruption
A single exception lands in the queue and triggers the orchestrator.
Main Operations Agent
The orchestrator receives the exception and fans work out across specialized sub-agents.
Five sub-agents, one decision
Each runs in parallel with a narrow job. Every step is shown.
Gathers live inventory, the affected orders, supplier track record, and DC capacity by querying NetSuite, Fishbowl, and Netstock.
Context built: 340 units short across 12 customer orders, DC-1 safety stock down to 15%, a 420-unit surplus sitting at DC-3, and backup supplier Lakeshore with room to ship.
Sorts the exception against 47 catalogued types and assigns a severity level.
Tagged as Type 12: Supplier Delay / Partial Shortfall. Severity: High. Fits a known, repeatable pattern (1 of 31 automatable types).
Works out the knock-on effect on orders, revenue, and customer commitments.
Revenue exposure: $21K across 12 orders. 3 of them carry delivery promises inside 48 hours, 2 key accounts are in scope, and alternative sourcing is on hand.
Weighs the resolution paths: reroute, expedite, substitute, or escalate.
Recommendation: move 280 units from DC-3 safety stock and expedite 60 from Lakeshore Parts for $1,400. The alternative, waiting three days, puts $21K of revenue at risk. Approve the reroute.
Builds the transfer orders, PO amendments, and supplier notices ready to fire.
Draft actions staged: a DC-3→DC-1 transfer order, a Lakeshore Parts expedite PO, a Pacific Components delay notice, and customer ETA updates. Held for approval.
How it learns
When a human corrects the agent, the rule updates itself.
The world moves; the rules follow
The system watches for changes in the world and rewrites its own rules.