Adyen Turns a 45-Minute Dependency Hunt Into Seconds With Architect

Adyen has built Architect, a near real time map of its platform that connects code, runtime behavior and infrastructure. The company says the system reduced a dependency investigation that once took 45 minutes to a few seconds for engineers and AI agents.
- Adyen says Architect cut a 45-minute investigation to a few seconds.
- The system combines service, code and infrastructure graphs.
- Its service graph processes up to 7 million tracing spans per second.
- A Neo4j composite database lets queries reach all three graphs.
- Adyen uses Architect for engineering, resilience and AI agent workflows.
A platform too large to map manually
Adyen provides payments, data insights and financial products through a single global platform. Its engineering team describes Architect as a response to a recurring production question: “What actually depends on this thing?”
Before Architect, answering that question meant switching between traces and a code repository. The necessary information could be spread across the code that makes a call, runtime traces that record it and the hosts where it runs. These areas were handled in separate tools by separate teams.
The scale made manual reasoning difficult. Adyen’s platform includes roughly 250,000 classes, around 3,000 live services and tens of thousands of hosts. Production can generate up to 7 million tracing spans per second. The team wrote that AI agents faced the same problem as people when large amounts of raw context were supplied: “Most workflows (for both agents and humans) share the same failure mode: they’re guessing.”
Three graphs connect code, services and hosts
Architect builds three separate graphs from how Adyen’s platform operates. The service graph covers around 3,000 live services and approximately 470,000 calls between them. Trace data from every workload and environment is sent to Apache Kafka, while Apache Flink processes the traces and extracts the service relationships.
The code graph covers roughly 250,000 classes across more than 6,000 modules. Rather than repeatedly rebuilding everything, Adyen uses incremental builds triggered by a git poller. The infrastructure graph covers tens of thousands of hosts across tens of data centers, with around 900 metadata items per host. Because infrastructure changes less frequently, scheduled jobs refresh that graph a couple of times per day.
Adyen stores the graphs in separate Neo4j databases and connects them through a composite database. This lets a single Cypher query reach several graphs without creating a merged copy, while each graph retains its own update pipeline and refresh rate.
A factual foundation for engineers and agents
In Architect, relationships directly connect services with the code behind them and the hosts on which they run. Adyen says one graph query can return what would otherwise be thousands of lines of context for an agent. The company describes the result as “a factual substrate: pre-computed, verified, and queryable.”
Adyen first built Architect for engineers, then gave AI agents access to the same map. The system is used for questions involving multiple domains, such as checking whether a runtime dependency is referenced in code, assessing which critical flows could be affected by a hurricane at a data center, or evaluating whether a service can be modularised without breaking a production caller.
The company has also implemented resilience health rules using the combined data. Other uses include security tabletops, disaster simulation and anti-pattern monitoring across code, runtime and infrastructure. Adyen is publishing further accounts of how agents use Architect for runtime incidents, compile-time complexity and infrastructure resilience, while describing the system as still under development.
Neo4j: A 45-Minute Answer to One Question


