4 · Production agentsAtlas · Control tower (broader)

Atlas · Control Tower Agent

At a glance

  • Live at: Multiple LSPs (control tower deployments)
  • Channels: API-driven (autonomous monitoring), voice (outbound calls to drivers/customers)
  • Languages: EN, HI (locale-dependent)
  • Owner: AI pod
  • Key stat: Processes incident tickets automatically — vehicle stationary, suspicious delivery patterns, delayed consignments

The problem it solves

A logistics control tower monitors thousands of shipments in real time. When something goes wrong — a vehicle stops moving, a delivery takes suspiciously long, a consignment misses its SLA — someone needs to investigate and act. At scale, human operators can’t triage fast enough. Incidents pile up, response times slip, and preventable losses grow.

Atlas automates the triage-investigate-act loop. It picks up incident tickets, analyzes the situation, contacts the relevant parties (drivers, dispatchers, customers), and takes action — all without a human in the loop unless the situation is genuinely complex.

How it works

Step by step

  1. Ticket intake — Incident ticket arrives from LIA (Query Builder) or ProjectX monitoring.
  2. Supervisor analysis — The control_tower_supervisor agent reads the ticket, identifies the incident type, and decides on the action plan.
  3. Outbound action — Routes to the appropriate voice agent:
    • Driver agent — calls drivers about trips (vehicle stationary, delays)
    • Dispatch agent — coordinates with dispatch for reallocation
    • Customer agent — verifies delivery, confirms receipt
  4. Ticket update — Posts findings as comments on the LIA ticket, updates status.
  5. Follow-up — If unresolved, schedules a follow-up call/action using the schedule_followup tool.
  6. Close — Marks the ticket as resolved when the issue is cleared.

Agent templates in the codebase

TemplateIDTypeModelRole
control_tower_supervisor2005sub_agentGemini 2.5 FlashProcesses incident tickets, decides action plan
control_tower_driver3002voice_agentGemini 2.5 FlashCalls drivers about trips (via ElevenLabs)
control_tower_dispatch3003voice_agentGemini 2.5 FlashDispatch coordination calls
control_tower_customer3004voice_agentGemini 2.5 FlashCustomer delivery update calls

Tools used

ToolWhat it does
fetch_informationFetch trip/consignment details from ProjectX
lia_post_commentPost findings as comments on the incident ticket
lia_update_ticket_statusUpdate ticket status (acknowledged → actionTaken → closed)
lia_update_ticket_custom_fieldUpdate custom fields with agent findings
schedule_followupSchedule follow-up calls/messages
object_actionGeneric ProjectX object actions (fetch, update)

Atlas vs Maya

AtlasMaya
ScopeGeneral-purpose control tower — multiple incident typesBDO-specific — card delivery monitoring
DeploymentMultiple LSPsBDO Unibank
Incident typesVehicle stationary, suspicious delivery, delays, address issuesCard delivery SLA, delivery confirmation, exception handling
CustomizationGeneric templates, configured per deploymentDeeply customized for BDO’s card delivery workflow

Atlas is the productized version of the control tower pattern. Maya was the first deployment (BDO-specific); Atlas generalizes it for any LSP.

Tech stack

LayerWhat’s used
ModelsGemini 2.5 Flash (cost-optimized for high-volume incident processing)
MemoryShort-term: conversation + ticket context. Long-term: incident resolution history.
Tools (MCP)ProjectX fetch, LIA ticket management, follow-up scheduling.
GuardrailsContent safety on voice calls. PII masking in ticket comments.
EvalsScenario-based per incident type; escalation rate tracking; follow-up effectiveness.

Sources

Changelog

  • 26 May 2026: Full content from GitHub repo analysis (control tower templates) and video library.