AI Agents ·AgentFleet

Astra — The Planning Agent

Autonomous routing, sequencing, and allocation — the thinking layer of AgentFleet.

✦  75%+ reduction in planning time and a 20% driver productivity lift at Southeast Asia's largest air-logistics network.

The problem it solves

Every morning, a dispatch supervisor looks at a shipment manifest, a driver roster, a map of the depot’s territory, and tries to answer four questions at once: Who goes where? In what sequence? In which vehicle? With what window promise? The answer changes by the minute. A cancellation at 7:42. A traffic incident at 8:05. A reload at 8:19. A driver calling in sick at 8:31. By 9:00, the plan the team spent two hours building is already stale, and the “plan” becomes “whoever shouts loudest.” That is how depots leak 20 percent of driver productivity and miss half their time-window promises — not because the software shows the wrong numbers, but because no one has time to re-plan every time reality moves. Astra exists to absorb that re-planning load. It doesn’t replace the planner. It removes the part of the planner’s day that doesn’t require judgement.

What it is

Astra is Shipsy’s Planning Agent — the thinking layer inside AgentFleet that decides who gets which shipment, in what sequence, on what route, and which depot or hub absorbs which load. It sits between the order book and execution, continuously. Where traditional TMS optimizers run once per batch, Astra runs a rolling optimization against live constraints — vehicle capacity, driver hours-of-service, time windows, skill tags, cold-chain requirements, fatigue thresholds, SLA priority, parking-spot availability — and emits decisions, not suggestions. It writes the plan back into the dispatch system, publishes the driver app sequences, and schedules the next re-plan window. What’s new-age about it: Astra doesn’t “optimize” in a vacuum. It handshakes with Clara the moment a window is at risk, with Nexa the moment a trip is costed, and with Atlas the moment a re-route breaches a policy. It is the planner that never goes home.

Core capabilities

Capability What it does
Multi-constraint route solver Column-generation optimizer that simultaneously respects capacity, time windows, skill/vehicle-class tags, driver hours, break rules, and loading-dock slots. Produces feasible routes, not just shortest paths.
Micro-cluster sequencing Uses the Micro-Cluster Route Optimization layer to respect driver-validated tribal knowledge — which door to approach, where to park, which walking order is faster for a 12-stop apartment complex.
Dynamic re-planning Every cancellation, reload, ETA drift, or driver exception triggers a scoped re-solve — only the affected legs, not the whole route. A full depot re-plan finishes in under 90 seconds.
Territory balancing Detects overloaded vs under-loaded territories in real time, re-draws the soft boundaries, and reallocates surplus jobs to adjacent drivers without breaching customer window promises.
Allocation scoring Assigns each shipment to the right fleet type (own, 3PL, hyperlocal, floating) based on a scoring function — SLA tightness, cost-per-drop, carrier performance scorecard, capacity left today.
Driver fatigue awareness Reads the DFMP (Driver Fatigue Management Plan) ledger. Refuses to allocate a 10-hour shift to a driver who has flagged rest hours. Governance, not just optimization.
Appointment-delivery planning Treats appointment windows as hard constraints, not soft preferences. Pre-books the linehaul backwards from the customer slot. Flags infeasible commitments to the sales team before they’re promised.
Depot-level load planning Integrates with Smart Depot Operations — sequences pick-face waves, guides loading order, and publishes the reverse-order load plan to ground staff.
Hub-and-spoke network planning Line-haul trip creation, bag consolidation logic, hub-to-hub capacity balancing. Air leg aware (IndiGo cargo AWB sync) and sea-leg aware (inter-atoll routing for island networks).
Confidence-tiered actions Decisions below 70% confidence surface to a planner with a suggested action and a one-click approve. Above 85% — autonomous execution. Tier logic is per-customer configurable.
Plan-vs-actual telemetry Every plan Astra publishes is scored against the actual outcome — arrival, dwell, success, cost. The scoring feeds the next day’s solver. The agent learns the depot.
What-if simulation Planners can ask: “What happens if I add two vans on Thursday?” or “What if this customer moves to a two-hour window?” Astra runs the counterfactual in under a minute.

How it works

Astra’s architecture is an event-driven loop. Orders enter the order book and emit a shipment.created event. The Constraint Assembler hydrates each shipment with its hard and soft constraints — window, capacity, vehicle class, skill, cold-chain, appointment ID, hazmat flag. The Solver Orchestrator decides whether this is an additive re-plan (add the new shipment to an existing route), a scoped re-solve (the shipment breaks an existing route’s feasibility), or a full wave re-plan (the change is large enough to rebalance territories). Each solver type has its own service — column-generation for dense urban, insertion heuristics for sparse rural, VRPPD for pickup-and-delivery. The Decision Writer pushes the plan to the dispatch system, the driver apps, and the customer ETA service. The Feedback Collector listens for every arrival, every dwell, every exception, and writes outcomes back into the training ledger.

graph TB A[Order Book] -->|shipment.created| B[Constraint Assembler] B --> C{Solver Orchestrator} C -->|additive| D[Insertion Solver] C -->|scoped| E[Column-Generation Solver] C -->|wave| F[Territory Rebalancer] D --> G[Decision Writer] E --> G F --> G G --> H[Dispatch System] G --> I[Driver App Sequences] G --> J[Customer ETA Service] K[Feedback Collector] --> L[Training Ledger] H --> K I --> K J --> K L -.feeds tomorrow.-> C

The execution loop is where Astra earns the word “agent.” It doesn’t just plan — it re-plans, negotiates with sister agents, and escalates when a decision crosses a policy line.

sequenceDiagram participant OMS as Order Book participant AST as Astra participant CLA as Clara participant NEX as Nexa participant ATL as Atlas participant DRV as Driver App OMS->>AST: shipment.created (tight window) AST->>AST: scoped re-solve (existing route) AST->>DRV: updated sequence Note over AST: ETA drift detected at stop 4 AST->>AST: recompute downstream stops AST->>CLA: slot-change risk on stops 5-7 CLA->>CLA: proactive SMS/WhatsApp to customers AST->>ATL: re-plan breaches cost policy ATL-->>AST: approved (tier 2) AST->>NEX: trip cost payload finalized NEX->>NEX: pre-audit against contract rate card

Proven outcomes

Customer type and scale Outcome
Southeast Asia’s largest air-logistics network, 164 hubs, 207 fleets 75%+ reduction in planning time; 20% driver productivity increase; 20+ hours/week ops savings
Leading Western European parcel operator, 50%+ national market share, 8M+ shipments 90%+ delivery-window adherence (from 30%); 20%+ driver productivity increase; USD 5M+/year savings
A premium Indian B2B express network, 49 cities, 3,500+ pincodes 16–18% cost savings; 90%+ FADR (from ~75%); real-time air-leg visibility
A LATAM quick-commerce grocery operator, 400+ drivers, 30-min SLA 98.5%+ orders auto-allocated; 90%+ adherence to 30-min SLA; USD 168K/year resource savings
A leading Australian parcel operator, AUD 200–250M revenue 10–15% driver productivity lift; ~35% reduction in failed deliveries; roster rules enforced by DFMP

Integrations

Astra runs inside the AgentFleet substrate and inherits the integration bench:

  • ERP and OMS — Oracle ERP, SAP, Salesforce, Veeva, custom order books via REST and Kafka
  • Driver and fleet telemetry — Wialon, Wheelseye, native Shipsy Driver App, third-party mobile SDKs
  • Carrier APIs — 240+ carrier integrations (hyperlocal, 3PL, regional), with IndiGo cargo AWB sync for India air-leg and inter-atoll sea-route awareness for island networks
  • Warehouse and depot — Shipsy Hub Ops App, Smart Depot scanners, WMS sync for pick-face wave sequencing
  • Data platforms — Snowflake, BigQuery, Databricks; every plan decision is stream-exported for BI
  • Sister agents — Clara (CX), Nexa (settlement), Vera (dispute resolution), Atlas (control tower)

Deployment

Phase 1 — Discovery (weeks 1–2). Depot walks, order-book schema review, constraint inventory. We name the top ten hard constraints and the top five soft ones, per depot. Success criterion: a constraint document signed off by operations.

Phase 2 — Configuration (weeks 2–5). Solver profile tuning (urban dense vs rural sparse vs mixed), driver-validated micro-cluster ingestion, territory bootstrap. Rate-card and SLA policies loaded into the Decision Writer. A shadow plan runs nightly alongside the incumbent planner.

Phase 3 — Pilot (weeks 4–6). One depot or region goes live. Human-in-the-loop gating stays on for the first ten days, then Tier 2 autonomy opens for additive re-plans. Planners keep full override authority.

Phase 4 — Scale (weeks 8–12). Progressive depot rollout, typically two to four depots per week. Tier 3 autonomy unlocks for wave re-plans where plan-vs-actual scoring is within tolerance. Governance dashboard is owned by ops leadership.

Most enterprises reach steady-state in 8–12 weeks, with measurable productivity lift in weeks 4–6. Success criteria are pre-agreed: planning time, window adherence, cost-per-drop, driver productivity. Governance is reviewed weekly in the first quarter, monthly thereafter.

Security and compliance

  • SOC 2 Type II, ISO 27001, GDPR-ready data handling
  • 21 CFR Part 11, GDP, and GMP-aligned audit trails for pharma and cold-chain deployments
  • Full audit trail on every plan decision — inputs, constraints, solver version, confidence tier, approver
  • Three-tier confidence scoring with configurable autonomy thresholds per customer and per depot
  • Human-in-the-loop gating for high-risk actions (large wave re-plans, cross-territory reallocations, policy-breaching cost decisions)
  • Data residency options — EU, India, Middle East, APAC regional hosting

Case study callouts

Southeast Asia’s largest air-logistics network · 164 hubs across APAC

Cut planning time by 75% and lifted driver productivity by 20%. Ops teams recovered 20+ hours a week previously lost to manual trip creation and rider allocation. A single dashboard now orchestrates orders, rider performance, and weekly business reviews across 164 hubs.

Read the full case study

Leading Western European parcel operator · 50%+ market share

Went from 30% delivery-window adherence to 90%+, lifted driver productivity 20%, and saved USD 5M+ per year — by encoding 20 years of courier tribal knowledge into a micro-cluster route solver and letting Astra re-plan the day as reality moves.

Read the full case study

A premium Indian B2B express network · 49 cities, 3,500+ pincodes

Greenfield B2B network built to UPS-grade SLAs from day one. 16–18% cost savings, 90%+ first-attempt delivery rate, 100% e-way bill and reconciliation coverage — driven by Astra planning on top of live IndiGo air-leg telemetry.

Read the full case study

Frequently Asked Questions

How long does deployment typically take?

Most customers go live within 8–12 weeks with a pilot depot or region in weeks 4–6. Value is usually visible inside 30–45 days — planning-time reduction and driver productivity move first. Full network scale follows in 90 days, with governance reviews weekly in the first quarter.

How does Astra handle last-minute disruptions?

Every event — cancellation, reload, driver exception, traffic drift, ETA slip — emits a signal. Astra decides between additive re-plan, scoped re-solve, or wave re-plan based on the blast radius. A full depot re-plan finishes in under 90 seconds. The driver sees a refreshed sequence; the customer sees a refreshed ETA.

Does it replace our planners?

No. It absorbs the repetitive re-planning work that burns out planners — the 40th reshuffle of the morning. Planners stay in command of exceptions, policy overrides, and customer escalations. In practice, a planning team of ten handles 3x the volume without adding headcount.

How does Astra respect driver fatigue and compliance?

Astra reads the DFMP (Driver Fatigue Management Plan) ledger as a hard constraint. It refuses allocations that breach rest hours, max shift length, or mandated breaks. For Australian and European operators, this is non-negotiable; for every operator, it's how you keep good drivers.

What happens when Astra is wrong?

Every decision carries a confidence score. Below a configured threshold, decisions surface to a planner with a one-click approve. Above it, Astra executes autonomously. The plan-vs-actual ledger scores every decision against the real outcome — dwell, cost, success — and feeds tomorrow's solver. Systematic error patterns trigger a model recalibration.

How does Astra work with our existing TMS?

Astra can run as the planning brain inside Shipsy's TMS or as a sidecar to an incumbent TMS via REST and Kafka. In sidecar mode, Astra reads the order book, publishes plans back to the TMS, and handshakes with the driver apps and dispatch screens. Most customers collapse to the Shipsy stack within a year, once they see the agent-to-agent handoffs with Clara, Nexa, and Atlas.

Does it work for quick-commerce and 30-minute SLAs?

Yes. The same solver runs on a different cadence — rolling sub-minute re-plans instead of wave re-plans — and the allocation scoring weights SLA tightness above cost. A LATAM quick-commerce operator auto-allocates 98.5%+ of orders and holds 90%+ adherence to a 30-minute promise on mixed fleets.

What's the autonomy model — how much control do we keep?

Three tiers, configurable per customer and per depot. Tier 1 — suggest only, planner confirms. Tier 2 — autonomous for additive re-plans, planner confirms for wave re-plans. Tier 3 — autonomous end-to-end with policy guardrails. Most customers start at Tier 1, move to Tier 2 by week 6, and unlock Tier 3 by week 12 for specific depots. Ops leadership owns the tier map.

How does Astra handle cold-chain and pharma constraints?

Temperature thresholds, excursion windows, and licensed-driver skill tags are treated as hard constraints. Astra will refuse to allocate a cold-chain shipment to a vehicle without an active temperature logger or to a driver without the required training tag. For pharma customers, the solver handshakes with the LIMS and Veeva QMS event streams — a batch-release delay flips a shipment's readiness tag, which re-enters the solver the moment it clears.

How does the plan get to the driver?

Sequences, stop-level instructions, parking-spot hints, and walking-pattern guidance are published to the native Shipsy Driver App or a third-party app via REST. Drivers see an ordered list; supervisors see the reverse-order load plan; the back-end keeps the plan, actuals, and ETA stream in a single ledger.

What's the data footprint and where does it live?

Astra's working memory is the live order book plus 30 days of plan-vs-actual history per depot. Data residency options are EU, India, Middle East, APAC, and customer-VPC deployments. Every decision is stream-exported to the customer's data platform so analytics teams work on the same source of truth as the solver.