Your Agent's Tool Calls Will Run Twice: Idempotency and Evaluation Gates in Java Services
Why every LLM tool call eventually runs twice, and how idempotency keys plus evaluation gates make that safe in Java services.
These articles are written and scheduled. Each one goes live on the date shown.
Why every LLM tool call eventually runs twice, and how idempotency keys plus evaluation gates make that safe in Java services.
Since Java 25, the AOT cache holds method profiles. They shorten the warmup – and cost speed in the long run. Measured and explained.
How an idempotent Spring Boot starter evolved from 0.1 to 1.1: exactly-once JDBC, production fixes, a stability promise, and a first outside contribution.
A coding assistant was given a real feature request against the execution engine of an open-source orchestrator. Three pull requests later, the code was plausible and still wrong. The author, who maintains that engine, explains why: the assistant had a generic world model of executors and a grep-built model of the code, but not the human one that holds the boundaries and the reasons. Then he hands it that model, and it works.
Build a full-stack web app using only Java: one Maven command, a table backed by Spring Data and a validated form in 5 classes, no JavaScript.
Constructor expressions, interface projections and generated read models: what each costs when a list screen needs paging and optional filters.
Notes from Porto TechHub 2026: Embabel and agentic AI planning, a comparison of Spring AI, LangChain4j, and Embabel, and the new defaults in Git 3.
Small, distroless, signed, and scanned clean, yet vulnerable? Six Java container security myths taken apart and gaps closed.
Codename One now targets embedded Linux: a touch UI that needs a twentieth of a Chromium kiosk's memory and no GPU, and a headless backend that served more...
Nobody knows everything about, well, anything. What matters is why each gap is there, what the rest of the team already covers, and whether anyone is allowed to say they don't know.
Using MicroProfile Reactive Messaging with Quarkus, Apache Kafka and virtual threads.
Your agent passed its unit tests. Is it any good? Learn how to use harness-evals to evaluate Java AI agents.
Libán Bande González on his book Skinning JavaFX Applications, the Vitra and Celestia libraries, and why CSS alone is not enough for modern JavaFX interfaces.
Recursive CTEs hit an O(b^D) wall, and no amount of indexing or tuning moves it.
ProcessBuilder and FFmpeg on JDK 21+: the stderr pipe deadlock, ffprobe JSON into records, live progress, clean cancellation and parallel encodes.
Why the usual if-else feature toggle turns into debt, how to design toggles cleanly in Java, and how Flipjack moves the decision out of your code.
A YAML domain model that regenerates the Spring Boot code when the architecture or stack changes, keeps AI-written code consistent, and exports an Apache Ossie model for BI.
JEP 500 makes final mean final. How Tentackle's TRIP serializer stopped writing final fields for good, and a blueprint you can reuse for your own framework.
A 148 ms closed-loop max, the samples wrk added around it, and a run beside a second JVM that read 2.4 s instead of 2 ms and still passed the gate.
A Spring Boot starter that budgets Spring AI ChatClient calls by tokens and cost, with per-call metering, atomic reservations, Redis support, and agent call limits.
A worker that dies after AGREE never sends its result. How the manager handles expired assignments, and what happens when a late response turns up anyway.
An introduction to OpenID AuthZEN, the standard API between PEPs and PDPs, with Keycloak as the PDP and authzen-enforcer, a Java SDK.
Why a JDK 25 AOT cache can crash with SIGILL on another CPU, and how Axelix builds portable multi-arch Docker images without cached adapters.
Compression put 31× fewer bytes on the wire and made the request 27 seconds slower.
From a Spring Boot app to a working CRUD with search, filters and validation, declared in Java only: how Mateu 3.0 does it and where it fits.
Code accretes because reality keeps delivering requirements nobody wrote down. The damage from this comes when the code is the only record of why, and the way out is to recover and record that specification.
An in-memory graph is fast until you need to write to it.
When twenty operations run billions of times and none of them is hot, a flame graph tells you about HashMap and nothing you can act on.