Engines
At a glance
Pre-built deterministic capabilities that agents call as tools. Unlike LLM-powered reasoning, engines are configure-once, deterministic, fast. They handle tasks where you want consistent, rule-based behavior — not probabilistic generation.
Why this matters
Engines are the backbone behind many agent capabilities. When an agent “optimizes a route” or “resolves an address,” it’s calling an engine — not asking the LLM to figure it out. Understanding which engines exist and what they do helps you scope deployments and explain to customers what’s deterministic vs what’s AI-powered.
Engine inventory
| Engine | What it does | Used by |
|---|---|---|
| Address Intelligence (Atlas) | Validates, normalizes, geocodes addresses. Maintains a resolution history that improves over time. | Address Checker agent, Vera, any delivery agent |
| Auto-Allocation | Assigns orders to carriers/vehicles based on rules (cost, SLA, capacity, zone). | Dispatch workflows, control tower |
| Alert Management | Generates and routes alerts when conditions are met (SLA breach, vehicle stationary, delivery exception). | Atlas/Maya control tower agents |
| Geocoding | Converts addresses to coordinates and vice versa. Supports Google Maps, internal geocoder. | Address Intelligence, route optimization |
| LIA (Query Builder) | Ticket management system. Creates, updates, and routes incident tickets. Agents post comments and update status. | Atlas, Maya, all control tower agents |
| Route Optimization | Calculates optimal delivery routes considering time windows, vehicle capacity, traffic, priority. | Dispatch, appointment scheduling |
| Territory Planning | Defines delivery territories/zones. Assigns addresses to zones for routing and allocation. | Auto-allocation, route optimization |
Engines vs LLM agents
| Engines | LLM agents | |
|---|---|---|
| Behavior | Deterministic — same input always produces same output | Probabilistic — responses vary |
| Speed | Milliseconds | Seconds (LLM inference) |
| Cost | No per-call token cost | Token-based pricing |
| When to use | Rules are known, behavior must be consistent | Judgment required, natural language understanding needed |
| Configuration | Rule sets, thresholds, weights | Prompts, tools, policies |
In practice, agents use both: the LLM reasons about what to do, then calls an engine to do it. For example, Atlas reads an incident ticket (LLM reasoning), decides the address needs fixing (judgment), then calls Address Intelligence (deterministic engine) to geocode and validate it.
How agents call engines
Engines are exposed to agents as tools in the tool registry. The agent’s LLM decides when to call which tool; the engine does the heavy lifting.
Configuration patterns
Each engine is configured per-customer:
- Address Intelligence: Country-specific rules, confidence thresholds, fallback providers
- Auto-Allocation: Carrier priority matrix, cost weights, SLA rules, zone mapping
- Alert Management: Alert trigger conditions, severity levels, routing rules
- Route Optimization: Vehicle profiles, time windows, traffic models, constraint weights
Configuration is done in the TMS/WMS console or via API. Once configured, agents inherit the settings — no per-agent engine setup needed.
Sources
- See Tools & MCP for how tools connect agents to engines
- See TMS & WMS consoles for engine configuration in the console
- See Address Checker for Address Intelligence in action
Changelog
- 26 May 2026: Full content with engine inventory, comparison matrix, and configuration patterns.