CHAPTER 03 · BUILD
Agents Shipped Inside the Existing Stack
The first system, and the two right behind it, shipped inside the terminal network's existing systems — no new TMS, no rip-and-replace. Each one is real and running.
THE AGENTS · 3 FOR THIS WORKFLOW
This load plan needed five agents working together — route, stops, fuel, hours-of-service, and a rest plan — because the work splits into five distinct jobs. Each is its own identity with its own scoped tools. Another workflow might need two.
Per-Load Route & Fuel Plan
Travel Center Demand & Workforce
Driver Retention Signals
IT DOESN'T STOP AT A REPORT
The Agent Takes the Action
A plan is one output. Most of what these agents do is take the next action across dispatch — a person approves anything that commits a truck or a carrier.
Build the load plan
Route, fuel stops, and an HOS-legal rest plan are planned together the moment a load is assigned.
Tender the load to the carrier
Sends the tender to the chosen carrier once dispatch approves it.
Send the customer the ETA
Pushes the updated arrival time to the customer automatically as the plan changes.
Book the terminal slot
Reserves the dock or appointment window against the planned arrival time.
PRODUCTION TRACE
Per-Load Route & Fuel Plan
Every step below runs against the plant's real data — the exact sequence, in order.
41
Load plans generated
36
Plans accepted as-is
3
HOS conflicts caught before dispatch
2
Fuel-stop overrides
3 items need your judgment
Ingest
AutomatedLoad details, driver HOS clock, and network fuel prices pulled the moment a load is assigned — no new data entry for dispatch.
Five-agent plan
AutomatedRoute, travel-center stop selection, fueling strategy, and an HOS-compliant rest plan are built together, not as four separate decisions that fight each other.
Ground every claim
AutomatedEvery recommendation cites the exact HOS clock, fuel price, or network constraint behind it, not a general opinion.
Dispatcher review
Reviewer inputThe full plan is one-click accept or send-back. Send-backs become training signal for the next load.
Load tracking
AutomatedPlan vs. actual gets tracked through delivery — did the stop, fuel, and rest plan hold up.
REAL OUTPUT · TRAVEL CENTER DEMAND & WORKFORCE
One Example, Fully Processed
When a job calls for a written artifact, this is what the system produces — everything it finds, structured for the action it triggers next.
Terminal 12 — Daily Ops Brief
Travel center, I-40 corridor · Generated 4:15 AM for the 6:00 AM shift change · Local + network signals
Signals Flagged for Today's Shift
| Signal | Reading | Threshold | Zone | Confidence |
|---|---|---|---|---|
| Fuel-island demand forecast | +22% vs. yesterday | Staffing trigger at +15% | Fuel islands | 93% |
| Weekend staffing gap | 2 shifts short, Saturday PM | Min. 4 staff per shift | Restaurant & store | 97% |
| Parking-lot occupancy trend | Trending toward full by 9 PM | 90% capacity | Overnight lot | 88% |
| Energy-cost anomaly | Refrigeration load +14% | ±10% of baseline | Cold storage | 64% |
| Store restock trigger | 3 SKUs below reorder point | Standard reorder rule | Convenience store | 95% |
Shift Brief — for the Terminal Manager
Saturday's fuel-island demand is running well above forecast, and the restaurant/store schedule is two shifts short to cover it. The refrigeration energy anomaly is worth a look but isn't urgent yet. Everything else on today's board is a normal restock cadence.
Tough Calls — Shift-Change Briefing
WAVE 2 — IN BUILD
Driver Retention Signals — What's Next
Drivers rarely quit over one bad week — it's a pattern of missed home time, pay variance, and unassigned miles that dispatch can't see until the resignation call. By then, the fix is too late.
The agent proposes which intervention to prioritize for a driver showing risk signals; a rules engine checks it against pay policy and contract terms. Nothing ships on an LLM's judgment call about compensation — the policy check is the check.
Worked example the system evaluates today in testing
A driver cohort out of Terminal 6 shows three risk signals in the same week: two missed home-time windows, unassigned miles trending down, and a pay-variance flag. Four intervention options scored by projected retention-risk reduction:
WHAT CHANGED
Time returned
~95% of manual planning time
Dispatchers got the time back for exception handling, not for building routine plans by hand.
Operating complexity
A 20-minute manual build became a 90-second reviewed plan
Where this stands
First system live in 30 days
UP NEXT · TEAM VIEW