Industry: Logistics & Supply Chain    |    Region: Singapore (14-country operations)    |    Client Type: Enterprise 3PL ($2.8B annual freight)

About the Client

The client is a Singapore-headquartered third-party logistics provider managing $2.8 billion in annual freight volume across 14 countries in Asia-Pacific, the Middle East and Europe. They handle everything from full-container-load ocean freight to last-mile delivery, serving manufacturers, retailers and e-commerce companies.

Their operations center in Singapore runs 24/7, managing roughly 42,000 active shipments at any given time. The logistics coordinators are the backbone of the operation and they were drowning.

The Challenge: 38% Exception Rate and 4.7-Hour Resolution Time

In global logistics, “exceptions” are the norm, not the exception. 38% of the client’s shipments experienced at least one disruption during transit: port congestion (especially at Tanjung Priok, Colombo and Jebel Ali), customs holds (documentation errors, classification disputes), carrier schedule changes (vessel delays, blank sailings) and last-mile delivery failures.

Each exception required a logistics coordinator to investigate across 6+ systems: the TMS (Transport Management System), WMS (Warehouse Management System), carrier portals (each carrier has a different portal with different data formats), customs databases, customer CRM and email/messaging threads with carriers and customs brokers.

Average resolution time: 4.7 hours per exception. With approximately 16,000 exceptions per month, the company was burning S$6.2 million per year on a 45-person exception management team. On-time delivery rate had dropped to 91.4% and their largest customer (an electronics manufacturer) had put them on notice.

Dextra Labs’ Agent Architecture: Three AI Agents with MCP Connectors and OpenClaw Skills

Here is Dextra Labs‘ SYSTEM ARCHITECTURE for Autonomous Supply Chain Control Tower:

┌───────────────────────────────────────────────────────────────┐
│           REAL-TIME DATA INGESTION LAYER                    │
│  AIS Vessel Tracking • Port Authority APIs • Weather APIs   │
│  Carrier EDI Feeds • Customs Gateway • TMS Webhooks        │
│  → All via MCP Connectors → Apache Kafka → Event Store    │
└───────────────────────────────────────────────────────────────┘
                          │
                          ▼
┌───────────────────────────────────────────────────────────────┐
│   AGENT 1: DISRUPTION DETECTION AGENT                      │
│   Correlates events across data streams to detect           │
│   disruptions BEFORE they hit the shipment                  │
│   Output: exception_event (type, severity, affected SKUs)   │
└────────────────────────────────┬──────────────────────────────┘
                                │ (A2A: exception_event)
                                ▼
┌───────────────────────────────────────────────────────────────┐
│   AGENT 2: ROOT CAUSE & RESOLUTION AGENT                   │
│   RAG over 3 years of resolution history (48,000 cases)     │
│   RAG over carrier SLA databases (220+ carriers)            │
│   Output: recommended_action + confidence score             │
└────────────────────────────────┬──────────────────────────────┘
                                │ (A2A: recommended_action)
                                ▼
┌───────────────────────────────────────────────────────────────┐
│   AGENT 3: ACTION EXECUTION AGENT                          │
│   OpenClaw skills for carrier API integrations              │
│   MCP → TMS (rebook), Customs (amend), CRM (notify)       │
│   Auto-executes for confidence > 0.85                      │
│   Routes to human coordinator for confidence < 0.85        │
└───────────────────────────────────────────────────────────────┘

Agent 1: Disruption Detection AI Agent

This ai agent doesn’t wait for someone to report a problem. It proactively correlates events across multiple data streams to detect disruptions before they impact shipments. AIS vessel tracking data shows a container vessel is running 18 hours behind schedule. The ai agent cross-references this against all active shipments on that vessel, identifies affected containers, calculates the downstream impact (missed feeder connections, customs filing deadlines) and fires an exception event before the carrier even sends a notification.

MCP connectors: AIS vessel tracking (MarineTraffic API), port authority APIs (PSA Singapore, DP World), weather APIs (OpenWeatherMap maritime), carrier EDI feeds (12 major carriers), customs gateway (Singapore TradeNet, Indonesia INSW).

Agent 2: Root Cause and Resolution AI Agent

When an exception is detected, this AI agent determines what caused it and what to do about it. It runs RAG over two knowledge bases: 3 years of resolution history (48,000 resolved exceptions with full context, what went wrong, what was tried, what worked) and carrier SLA databases (220+ carrier contracts with penalty clauses, free-time allowances and rebooking policies).

The ai agent produces a recommended action with a confidence score. For a vessel delay causing a missed feeder connection, the recommendation might be: “Rebook on [Alternative Carrier] sailing [Date] from [Port]. SLA: carrier covers rebooking cost per clause 4.2 of MSA. Alternative adds 2 days to transit. Customer notification required.” Confidence: 0.92.

Agent 3: Action Execution AI Agent

For recommendations with confidence above 0.85, the Action Execution ai Agent auto-executes: rebooks with the alternative carrier via OpenClaw-based skills that interface with carrier booking APIs, files customs amendments via TradeNet MCP connector, updates the TMS and sends templated customer notifications via the CRM. For confidence below 0.85, it routes to a human coordinator with the full analysis pre-loaded.

We used OpenClaw’s skill architecture for the carrier API integrations because each of the 220+ carriers has a different booking API format. OpenClaw skills act as adapters, each carrier gets a skill that translates the standard rebooking request into that carrier’s specific API format. Adding a new carrier means writing one new skill, not modifying agent logic.

Dextra Labs’ Workflow: A Container Gets Stuck at Customs

Step 1: Disruption Detected

The Customs MCP connector receives a hold notification from Singapore TradeNet: Container MRSK-4821 held for HS code classification dispute. The Disruption Detection ai Agent identifies 3 SKUs in that container, determines the customer (electronics manufacturer) and calculates the impact: 2-day delay will cause a stockout at the customer’s assembly plant.

Step 2: Root Cause Analysis

The Root Cause ai Agent queries the resolution history RAG: “HS code disputes for electronic components in Singapore have been resolved 87% of the time by providing supplementary product datasheets. Average resolution: 6 hours with datasheet, 3.2 days without.” It also pulls the customer’s customs file and finds the missing datasheet.

Step 3: Resolution Recommendation

Recommended action: “Submit supplementary product datasheet (already on file in customer record CRM-8821) to TradeNet via customs broker. Simultaneously prepare alternative routing via air freight from Hong Kong warehouse as backup if customs hold exceeds 24 hours.” Confidence: 0.91.

Step 4: Auto-Execution

Confidence > 0.85, so the Action Execution ai Agent files the supplementary documentation via TradeNet, notifies the customs broker, alerts the customer with an ETA update and pre-books the backup air freight option (held for 24 hours without commitment).

Step 5: Resolution Tracking

The system monitors TradeNet for the hold release. When the container is cleared (8 hours later), it cancels the backup air freight and updates the TMS with the revised ETA. The customer receives an automated “resolved” notification.

Step 6: Learning

The full exception lifecycle, cause, recommendation, action, outcome, is added to the resolution history database. Next time an HS code dispute hits for electronic components, the system will recommend the datasheet submission even faster, with higher confidence.

Implementation Decisions

Decision Area	What We Chose	Why / Trade-Offs
Carrier integrations	OpenClaw skill architecture over custom API adapters	220+ carriers, each with a different API. OpenClaw skills provide a modular adapter pattern. Adding a new carrier: 2–4 hours to write a skill vs. 2–3 days for a custom adapter. Skills are version-pinned and audited per OpenClaw security best practices.
Resolution knowledge	RAG over historical exceptions + carrier SLAs over rules engine	A rules engine would need 500+ rules and constant maintenance. RAG over 48,000 historical resolutions captures institutional knowledge that no rules engine can encode. The system gets smarter with every resolved exception.
Auto-execution threshold	0.85 confidence for auto-execute, <0.85 for human routing	Set conservatively at launch. After 3 months, analysis showed 0.85 threshold was optimal, lowering to 0.80 increased auto-resolution by 8% but introduced a 2.1% error rate that caused costly misroutings.
Data integration	MCP connectors for all external systems over point-to-point integrations	MCP provides a standard protocol for 8+ external systems. When Singapore launched the enhanced TradeNet API in Q3, we updated one MCP connector. Point-to-point would have required updating every service that touches customs data.
Deployment	AWS ap-southeast-1 (Singapore)	Client HQ is in Singapore. Most critical shipping routes transit through Singapore. Lowest latency to PSA port systems and TradeNet. Compliant with Singapore PDPA.
Proactive detection	Cross-stream correlation over per-system alerts	Individual system alerts (carrier notification, port congestion report) are reactive and often late. Cross-stream correlation (vessel AIS + weather + port congestion + booking data) detects disruptions 4–8 hours earlier than carrier notifications.

Tech Stack

Claude 3.5 Sonnet  ·  OpenClaw (carrier skills)  ·  MCP connectors (8+)  ·  A2A Protocol  ·  Pinecone (RAG)  ·  Apache Kafka  ·  AWS ap-southeast-1  ·  PostgreSQL  ·  Python FastAPI  ·  Next.js (control tower UI)

Results

73% Exceptions resolved autonomously

4.7h → 22min Average exception resolution time

S$6.2M → S$1.8M Annual exception management cost

99.2% On-time delivery (was 91.4%)

73% of exceptions are now resolved without any human involvement. The remaining 27% are genuinely complex cases, multi-party disputes, force majeure events, regulatory changes, that appropriately require human judgment. The human team went from 45 people to 14, with the retained team handling complex cases and strategic carrier relationship management.

On-time delivery rate improved from 91.4% to 99.2%. The electronics manufacturer customer went from issuing a performance notice to renewing their contract for 3 years with a 15% volume increase. The control tower UI, built in Next.js, gives the remaining team real-time visibility into all 42,000 active shipments with the agent’s status for each.

“We went from fighting fires to preventing them. The proactive detection is what changed the game, we know about problems before our carriers do.” – VP of Operations, 3PL Operator, Singapore

Engagement Timeline: 13 Weeks

Weeks 1–3: discovery, data source mapping, carrier API audit (220+ carriers, prioritized top 40 by volume).

Weeks 4–7: Disruption Detection AI Agent, MCP connectors for AIS/port/weather/carrier/customs.

Weeks 5–9: Root Cause AI Agent, RAG pipeline over 48,000 historical exceptions.

Weeks 8–11: Action Execution AI Agent, OpenClaw skill development for top 40 carriers.

Weeks 12–13: control tower UI, parallel running, production deployment.