Value classes and faster memory allocation to agentic development, AI model advances and more resilient Spring AI applications.
🎯 Featured Article
🚀 Java and Kotlin news: value classes arrive in JDK 28 and Kotlin 2.5 Flatter memory and lower overhead — value classes for Integer and LocalDate. Kotlin 2.5 adds multi-property value classes.
- Removes object identity from Integer and LocalDate (JEP 401).
- Adds multi-property value classes in Kotlin 2.5.0-Beta2.
⏩ TL;DR (Quick Recap)
- Value classes in JDK 28 and Kotlin 2.5 for flatter memory layouts.
- Java 21 to 25 upgrades for failures that surface silently.
- Spring AI 2.1 message parts to preserve reasoning across requests.
- Agent sandboxes and keep core skills sharp as agents write code.
☕ JVM Corner
🔒 Java Secure Coding Guidelines Secure Coding Guidelines for Java, with insecure examples and updated input validation and serialization coverage.
- Pairs insecure code examples with fixes that avoid vulnerabilities.
- Expands coverage of input validation and serialization.
🧠 Allocating Off-Heap Memory: Typically 6–18x Faster in JDK 28 JDK 28 serves small confined-arena FFM allocations from native pools, cutting allocation latency by up to 90%.
- Applies to Windows, Mac and Linux without code changes.
- Preserves all safety guarantees of confined arenas.
⏱️ Java’s new AOT cache feature is in a rough state AOT cache improves startup and warmup through training runs.
- Records loaded classes and called methods during a training run.
- Skips initialization work by reusing loaded and linked classes.
⚠️ Free Java, Real Risk: Why Unsupported OpenJDK Is Now a Security Gamble Unsupported OpenJDK in production no longer justifies self-patching and its weeks of lag against current attack timelines.
🔄 Upgrading from Java 21 to 25: What Actually Breaks Run on JDK 21 and 25 show which removals break builds loudly and which fail silently.
🪶 A Smaller, Faster Java Server with a Spring-Style API A native Java backend offers Spring-style compiled into a small executable.
🛠️ A beginner’s guide to the Jenesis build tool Build tool whose engine is plain Java code embedded in the project’s build/jenesis folder.
🍃 Spring Updates
🔎 Spring AI Recipe: Speeding Up RAG with Faster Embeddings Benchmarks five embedding options and shows how a Spring AI configuration change speeds up RAG pipelines.
- Identifies the embedding model as a likely bottleneck.
- Compares five options through configuration changes.
🔁 Spring AI Recipe: Switching to the OpenAI Responses API Spring AI 2.1 switches to OpenAI’s Responses API with one configuration property.
- Switches APIs with a single configuration property in Spring AI 2.1.
- Unlocks capabilities beyond traditional chat completions.
🧩 Spring AI 2.1 Message Parts: Reasoning That Survives the Round Trip Spring AI 2.1 models messages as ordered lists of parts so reasoning, signatures and tool calls survive round trips.
- Maps provider blocks, parts, items and chunks to message parts.
- Prevents HTTP 400 errors and silent loss of earlier reasoning.
📊 Lightweight Spring Boot Monitoring Without Prometheus and Grafana Basic visibility for small Spring Boot deployments on a VPS without a full Prometheus and Grafana stack.
🔍 Extra Reads
Frontier Models
🤖 GPT-6.1 Terra — Fast model with stronger visual reasoning, computer use and clearer writing than GPT-5.6 Terra.
🇪🇺 Mistral Large 4 — Open-weight 1-trillion-parameter multimodal model with three reasoning modes, led by coding and security benchmarks.
Agentic Development
🤔 Why are good developers so wary of agentic development? — Explains why developers distrust agentic tools: built is not the same as working in production.
📝 Anthropic Just Revealed 12 New Rules for Prompting Claude Opus 5.5 — Prompting rules for Claude Opus 5.5 that remove unnecessary instructions and avoid slow, costly agents.
📦 Not All Agent Sandboxes Are Created Equally — Compares agent sandboxes across four axes: boundaries, environment, lifecycle and interface.
🗺️ Here Be Monsters: What Do You Still Need to Know When Agents Do the Work? — Argues that owning software will depend on knowing enough to detect problems and dive deep when needed.
Working
🧭 Two Years of AI Changed 20 Years of Habits — A personal account of how AI changed daily engineering habits at Atlassian.
🏋️ Getting good at AI takes time and practice — Argues that working with AI is a skill requiring time and practice, given its uneven strengths.
Originally posted on marconak-matej.medium.com.