Agentic AI demonstrations can be remarkably persuasive. Give a model a goal and a collection of tools, and it appears to plan, make decisions and complete a complex task autonomously. Moving from such a demonstration to dependable business software, however, exposes a less glamorous set of problems.
In their Devoxx Belgium talk “A Year of Agentic AI Evolution”, Mario Fusco and Kevin Dubois describe how LangChain4j’s agentic framework has evolved in response to real-world experience. The emphasis has shifted from a fixed catalogue of agent patterns towards flexible orchestration—and towards capabilities such as tracing, persistent state, failure recovery and human intervention.
These operational concerns are the missing half of many conversations about agentic AI.
Predefined patterns remain useful. A supervisor can delegate work to specialised agents, while sequential or parallel workflows can organise more complex tasks. Yet business processes rarely fit one universal pattern. Processing an order, assessing an insurance document and handling a support request each require different combinations of automation, validation and escalation. Production systems therefore need composable workflows rather than an opaque loop that simply runs until the model decides it is finished.
State is particularly important. An agent may call several services, wait for external information or pause for approval. Its progress must survive an application restart or temporary outage. A chat history alone is not sufficient: the system needs a durable record of completed steps, tool calls, intermediate results and pending actions. Otherwise, recovery means starting again—with the risk of sending the same email, creating the same order or charging the same customer twice.
Failure recovery follows naturally. Language models time out, APIs become unavailable and generated tool arguments can be invalid. Java developers already know the necessary techniques: explicit transaction boundaries, idempotent operations, controlled retries, compensating actions and observable error states. Agentic software does not make these principles obsolete. It makes them more important because execution is less predictable than in a conventional workflow.
Human control should likewise be designed as part of the process rather than added as an emergency switch. Before an irreversible or sensitive action, an agent can enter a persistent waiting state and present an employee with the relevant context, its proposed action and the reasons behind it. The employee can approve, amend or reject the proposal, after which the workflow continues without losing its previous work.
This is where Java and Spring remain highly relevant. Their strengths are not limited to connecting an application to a language model. The ecosystem already provides mature solutions for data persistence, security, messaging, transaction management, retries, metrics and integration with business systems. Those capabilities form the operational foundation around the probabilistic parts of an AI application.
The real measure of an agent is therefore not how autonomous it appears in a demonstration. It is whether the surrounding system can explain what happened, resume after a failure and return control to a person at the right moment. Once state, recovery and human oversight are treated as first-class concerns, agentic AI can become less of an experiment and more of a dependable component of business software.
One comment