Inside Locus: How Fireworks and DiSCO Turn 250+ Constraints Into a Delivery Plan

On a normal peak morning, a delivery network running a few thousand drivers out of dozens of hubs will make tens of thousands of small decisions before the first van leaves the yard: which driver gets which route, which carrier picks up the overflow, which hub absorbs a volume spike from a neighboring one, which customer gets a delivery-window change instead of a missed promise. In most transportation stacks, those decisions are split across five or six disconnected systems and a Slack channel, reconciled by planners who are, in effect, running the real optimization engine in their heads. Locus was built on the premise that this coordination problem, not the routing math alone, is what actually determines whether a delivery network hits its numbers. This piece is a technical look at the two components that do that work: the Fireworks routing engine, which solves the sequencing and assignment problem, and DiSCO, the eight-agent decision layer that coordinates everything routing touches but does not itself decide.

What Fireworks and DiSCO Actually Are

Fireworks is Locus’s routing and dispatch engine. It takes a day’s orders, whatever mix of deliveries, pickups, service visits or reverse-logistics tasks a network runs, and solves for the sequence, assignment and timing that satisfies every hard constraint in play while minimizing cost. It is built to process more than 100,000 routes simultaneously and to re-solve continuously as the day’s conditions change, not to produce a single static plan at the start of the shift.

DiSCO, short for Digital Supply Chain Officer, is the layer that sits around Fireworks and decides which decisions Fireworks should even be asked to solve. Routing is only one of the decisions a delivery network makes on a given day. DiSCO coordinates eight specialist agents, Capacity, Carrier, Dispatch, Hub, Customer, Settlement, Orchestrator and Copilot, each owning one category of operational decision and each running the same underlying discipline: sense the current state, decide what to do, execute it, and learn from the outcome. Dispatch is the agent that calls Fireworks. The other seven exist because a routing engine that does not know whether a carrier has capacity, whether a hub is about to miss its cutoff, or whether a customer needs to be told about a delay is solving only a fragment of the actual problem.

Core Capabilities

How the Decision Actually Gets Made

Step 1: Orders and tasks enter the system

Deliveries, pickups, service visits and reverse-logistics tasks are ingested with their addresses, promised windows, service-time requirements and any fulfillment rules attached, from whatever order source, OMS, e-commerce platform or TMS, generated them.

Step 2: The Capacity and Carrier agents establish what is actually available

Before a single route is drawn, the Capacity agent resolves what owned, contracted and gig capacity is realistically available for the day, and the Carrier agent checks what volume can be tendered out across the partner network, so Fireworks is solving against real constraints, not a static fleet list.

Step 3: Fireworks solves the routing and assignment problem

Against that resolved capacity, Fireworks sequences stops and assigns them to vehicles and drivers, evaluating the full constraint set, time windows, capacity, driver skills, hub cutoffs, live traffic, cost, in one pass rather than as separate sequential steps.

Step 4: The Hub agent checks facility-level feasibility

Each route is checked against the sorting and dispatch cutoff constraints of the hub it originates from, so a plan that looks efficient on the road is rejected if the hub cannot actually release it on schedule.

Step 5: The Customer agent commits and communicates the promise

Once a route is finalized, the Customer agent sets or updates the delivery promise shown to the end customer, keeping the commitment tied to the live plan rather than a static estimate generated earlier.

Step 6: Execution begins and the Orchestrator agent watches for exceptions

As drivers execute the plan, the Orchestrator agent monitors for disruption, a delay, a failed attempt, a new urgent order, and triggers the relevant agent, usually Dispatch, to re-solve the affected portion of the plan within the autonomy level configured for that decision type.

Step 7: Settlement and the learning loop close it out

Once a route completes, the Settlement agent reconciles the financial record against what was actually executed, and every agent’s outcome data, what was planned, what happened, what was corrected, feeds back in as a training signal for the next cycle, the same sense-decide-execute-learn discipline every agent runs on a continuous basis.

Governance: How Autonomy Stays Accountable

Agentic decisioning only works at enterprise scale if it is auditable, which is why every agent in DiSCO operates inside a six-mechanism governance layer rather than as an unconstrained black box.

Explainability means every autonomous decision can be traced back to the specific constraints and data that produced it, in language an operator, not just an engineer, can review.

Traceability keeps a full record of every decision and the context around it, so an exception, a dispute or an audit can be reconstructed after the fact rather than argued from memory.

Evaluation continuously scores agent decisions against actual operational outcomes, so drift or degradation in decision quality is caught rather than discovered months later in a KPI report.

Autonomy levels, set from L1 through L3, let an operator decide how much independent authority a given decision type gets: L1 for recommendation-only, escalating to full autonomous execution once a decision type has proven reliable.

Execution sandbox lets a new policy or a configuration change be tested against historical or simulated conditions before it touches a live operation.

Human-in-the-loop ensures a person can always override, approve or reconfigure any agent decision, with full visibility into why the system proposed what it proposed.

What This Looks Like Across Three Different Deployments

A Fortune 50 parcel network: from manual override to governed autonomy

Running Orchestrator and Dispatch across pickup, transit and delivery decisioning for more than 4,500 drivers, this enterprise had been managing exceptions manually across 51 service-center locations, with capacity that existed on paper but was not visible to the people making daily assignment calls. Moving that decisioning onto DiSCO, with every autonomous action logged for explainability and override, raised weekly execution rate from 75 percent to 92 percent and surfaced more than $14 million in annualized operational opportunity that had been sitting inside the network the whole time, uncounted because no single system had visibility into it.

See the full deployment write-up

A North American retailer: six systems, one decision layer

Operating a multi-hundred-store network across ocean, rail and road, this retailer had been running six separate legacy systems to plan and execute its network, with exceptions resolved by whichever team happened to own the system where the problem surfaced. Consolidating onto one platform cut manual dispatch decisions by more than 80 percent, reduced exception resolution time to under two hours, and delivered more than $1 million in savings with break-even inside the first year.

See the full deployment write-up

A beverage distributor: from an hour of Excel to a live dashboard

Running depot-based distribution across mixed fleets of vans, trucks and motorbikes serving thousands of small retail points a day, this operator had been spending more than an hour on manual route planning before a single vehicle moved, with no validated delivery locations and no single view of the fleet. Locus’s routing and dispatch engine cut fuel consumption by 37 percent, increased orders per delivery trip by 22 percent, and reduced route planning time by 35 percent.

See the full deployment write-up

Integrations

Fireworks and DiSCO connect to a delivery network’s existing systems rather than requiring them to be replaced. The platform is built API-first, with pre-built connectors into TMS, WMS, OMS and ERP systems, so order data, inventory state and fulfillment rules flow in without a custom integration project for each source system. On the carrier side, ShipFlex connects a network of more than 1,000 pre-integrated carrier partners, with 160-plus actively used across current deployments, so the Carrier agent tenders volume without a manual onboarding cycle for every new partner. Telematics and driver-app data feed in continuously to give the Capacity and Dispatch agents live visibility into vehicle location, driver status and execution progress, the same signal every agent’s sense-decide-execute-learn loop depends on.

Deployment: How a Network Actually Gets to Autonomous Operation

Locus does not treat go-live as the finish line, because an agentic system only earns higher autonomy levels once its decisions have proven reliable in a specific network’s conditions. Deployment runs in three phases. In months one and two, Forward Deployed Engineers stand up the platform against the network’s actual data, configuring constraints, integrations and initial autonomy levels, typically starting most decision types at recommendation-only. In months three through six, agent decisions are tuned against live outcomes and autonomy levels are raised for decision types that have proven out, moving specific workflows from human-approved to fully autonomous execution. From month six onward, the customer’s own operations team owns the system directly, with the Forward Deployed Engineers training themselves out of the day-to-day loop rather than remaining a permanent dependency.

Security and Compliance

An agentic system making autonomous operational decisions has to meet a higher bar for auditability and data protection than a reporting tool, which is why Locus’s certifications cover both operational and data-handling standards. The platform holds SOC 2 Type II and SOC 3 attestations, is compliant with ISO 27001:2022 and ISO 27701:2019, and meets GDPR and HIPAA requirements where applicable. Access is controlled through SAML and ADFS single sign-on, role-based access control with granular permissions, enforced password policies, and Zero Trust Network Access. Data is protected with AES 256-bit encryption at rest and TLS 1.2 in transit. Combined with the governance layer’s explainability and traceability mechanisms, this means every autonomous decision is both auditable after the fact and protected in transit and at rest before it is ever made.

Recognized by the Analysts Who Evaluate This Category

Locus has been recognized by Gartner for seven consecutive years, including inclusion in the 2026 Hype Cycle for Supply Chain Execution and Logistics Technologies. QKS Group named Locus a Leader in Transportation Management Systems in its SPARK Matrix. G2 ranked Locus #1 in Route Planning in its 2026 Best Software Awards. Across the deployments this piece describes and the broader base of 360-plus enterprise customers in 30-plus countries, Locus has optimized more than 1.5 billion deliveries and delivered more than $320 million in aggregate logistics cost savings.

In October 2025, Ingka Investments, the investment arm of Ingka Group, the world’s largest IKEA retailer, acquired Locus. Locus continues to operate independently.

FAQs

What is the difference between Fireworks and DiSCO? Fireworks is the routing and dispatch solver, the engine that sequences stops and assigns them to vehicles and drivers against a full constraint set. DiSCO is the broader eight-agent decision layer that coordinates capacity, carrier allocation, hub feasibility, customer communication and settlement around that routing decision, so Fireworks is solving against real, current conditions rather than a static input.

How many constraints does Fireworks actually evaluate? More than 250 real-world constraints per computation, including vehicle capacity, driver shift hours, delivery time windows, road and access restrictions, live traffic and cost, solved simultaneously rather than as separate sequential rules.

What does the sense-decide-execute-learn cycle mean in practice? Each DiSCO agent continuously senses the current operational state relevant to its domain, decides what action to take, executes that action, and feeds the outcome back in as a training signal for its next decision, rather than running on a fixed rule set that never updates.

How much autonomy do the agents actually have? That is configurable per decision type, from L1, recommendation-only, through to L3, full autonomous execution. Most deployments start conservative and raise autonomy for specific workflows only after those decisions have proven reliable against live outcomes.

Can a human override an agent’s decision? Yes. Human-in-the-loop is one of the six governance mechanisms built into every agent, and every autonomous decision is logged with the reasoning behind it so an operator can review, override or reconfigure it at any point.

How long does it take to get from deployment to autonomous operation? Typically six months on the standard three-phase model: two months to stand up the platform and integrations, three to four months to tune agent decisions against live outcomes and raise autonomy levels, and from month six the customer’s own team owns the system directly.