When most people imagine Maersk, they see ships, containers, and ports. But behind the physical movement of goods sits a deeply complex digital logistics fabric. Every booking, customs clearance, yard movement, inventory update, and customer status check happens through APIs.

Maersk’s developer platform supports:

• 3,500 developers
• 1,500+ APIs
• 4,000+ downstream consumers running critical supply chain flows 24X7

At this scale, failures aren’t hypothetical; they’re inevitable.

Maersk created Stargate, an operational AI agent that automates the full first-level triage process: detection, RCA, blast-radius analysis, and routing. It is not a chatbot. It's an orchestrated, event-driven, multi-agent system integrated deeply into Maersk’s developer platform and observability stack. And it runs in seconds.

1. Autonomous detection powered by observability

Stargate begins with event-driven detection. Whenever an API is onboarded, Maersk’s platform automatically:

• Configures gateway routes
• Registers the service in the mesh
• Sets up Akamai configs
• Creates alerting rules in Grafana and Prometheus

So when an API misbehaves, Grafana fires an alert to a Webhook that the agent listens to. This Webhook trigger initiates the entire automated investigation.

2. Orchestration engine that gathers the ground truth

Once triggered, the Stargate orchestrator:

• Fetches proxy metadata (proxy ID, API name, mapping)
• Queries service mesh health
• Retrieves backend application status
• Looks up ownership and on-call teams
• Identifies consumers of the failing API

This data comes from Maersk’s internal metadata APIs which themselves talk to:

• Kong Admin APIs
• QMA service mesh
• Maersk’s developer platform
• Hedwig (on-call management platform)

“An agent is only as good as its integrations,” Das said. This integration layer is what gives the agent enterprise-grade context.

3. Two-agent RCA model: Log Analyzer + RCA Synthesizer

Stargate uses a two-agent architecture.

Agent 1: Log Analyzer

Logs contain more than 80 attributes — too many to send directly to the LLM. So before analysis:

• Sensitive data is stripped out (keys, IPs, etc.)
• Only relevant attributes are extracted, such as timestamp, trace IDs, status codes, API path, total request time, total proxy time, total target time

These custom metrics are crucial:

• total_target_time = 0 means request never reached the backend
• proxy time vs. target time reveals where latency is introduced
• error messages indicate DNS or mesh failures

The team explicitly instructs the model on what each attribute means. This dramatically increases accuracy. Providing definitions and examples, a technique called few-shot learning, improved agent accuracy from ~50% to ~80%.

Agent 2: RCA Agent

With the structured log analysis and metadata, the RCA agent performs:

• Pattern recognition across errors
• Correlation with mesh health
• Hypothesis generation (e.g., DNS failure, backend down, policy failure)

Its goal isn’t to provide a final fix, it’s to provide the correct direction within seconds.

4. Automated blast-radius analysis

One of Stargate’s most business-critical capabilities is blast-radius detection. Using metadata APIs, the agent identifies how many consumers use the failing proxy, which teams are affected, and which business flows may be impacted.

The agent automatically detects the kind of insights that used to take multiple people 15–20 minutes to establish, including number of consumers affected, downstream business impact, no authentication/security issues, and if the root cause is in the service mesh, not the gateway

5. Instant notifications with actionable context

Stargate pushess insights to Teams channels (Slack in future). A typical output includes:

• Summary of what failed
• Proxy ID
• Gateway link
• Access logs
• Application health
• Mesh health
• Blast-radius details
• Hypothesis of the root cause
• Links to start debugging immediately

The agent’s summary arrives only a few seconds after the alert fires — before most engineers even join the bridge.

“With this information in hand, I can join the bridge confidently,” Kumar said. “In our demo scenario, gateway wasn't the problem at all, and we knew that instantly.”

6. Built for future autonomy

Stargate was designed as an operational automation framework, not just a tool. It's intentionally modular, and future phases may include:

• Automatically removing unhealthy data planes from the load balancer
• Rerouting traffic to healthy regions
• Triggering owner notifications via Hedwig without human involvement
• Applying automated mitigations based on policy

Maersk’s deployment of Stargate transformed incident management. 

First, it impacted time, helping the team go from 20 minutes of triage to seconds.

Before Kong, incident response relied on lengthy bridge calls involving multiple teams. Engineers spent significant time manually digging through logs, and triage was often based on educated guesses about the failure's origin. With Kong, however, logs are analyzed instantly, and hypotheses are generated automatically. The blast-radius of the incident is calculated in seconds, allowing engineers to join the bridge with immediate, complete situational awareness.

Second, Stargate has taken over the draining, repetitive tasks that typically fatigue on-call personnel, including fetching logs and identifying proxy IDs, checking the health of the mesh and backend systems, evaluating DNS and service discovery mechanisms, mapping system owners, and formulating initial hypotheses for an issue. As a result, platform teams can now focus their valuable time on resolving issues rather than merely discovering them.

It's also improved the accuracy and consistency of triage. By structuring and explaining attributes like total_target_time, the agent interprets logs the same way senior engineers do but faster. This means more accurate differentiation between gateway vs. mesh vs. backend failures, less finger-pointing during incidents, and faster routing to correct teams.

Stargate has also resulted in clearer business visibility via blast-radius detection. Business stakeholders gain immediate visibility into the entities impacted, the volume of affected consumers, the potential for SLAs to be compromised, and the potential for revenue jeopardy, which facilitates more prompt customer communication and proactive risk mitigation.

Finally, Stargate has led to a cultural shift toward shared ownership. Embedding incident awareness directly into automation led to several positive shifts, including empowered developers who feel more capable and in control, reduced dependency on specific individual experts, enhanced visibility for teams into ongoing issues, and a more strategic focus for platform teams who can dedicate their efforts to higher-level improvements. Ultimately, Stargate did more than just streamline incident management; it fundamentally transformed how teams work together when facing high-pressure situations.

Maersk’s next decade of operational excellence

Stargate is already expanding into:

• Automatically modifying load-balancer pools
• Proactive data plane health management
• Intelligent alert threshold tuning
• Automated ownership routing via Hedwig
• Real-time anomaly detection powered by traffic patterns

Every improvement is evaluated through a consistent lens: does it empower engineers without compromising safety, governance, or operational resilience?

By combining Kong’s gateway, QMA mesh, their internal developer platform, and AI-driven automation, Maersk has built a modern operational fabric that scales not just its technology, but its people.