Viktoria Evdokimova on foojay.io - Friends Of OpenJDK

Can You Explain the Diff Your AI Agent Just Wrote?

Tue, 15 Sep 2026 14:12:01 +0000

Every Java developer has been in a code review where the question is simple: “What does this method do?”

The answer used to tell you something. The developer who wrote it would describe the intent, the edge cases, the trade-off they considered and rejected. The code review was a conversation about the invariant the code protects.

Now the answer is sometimes: “The agent wrote it.”

The code compiles. The tests pass. The linter is clean. Nobody in the review can explain the architecture of the PR, including the person who submitted it.

One Base URL, Many Coding Agents: What an OpenAI-Compatible Gateway Looks Like From java.net.http

Wed, 09 Sep 2026 16:09:53 +0000

Somewhere in every coding agent there is an HTTP client. Claude Code, Codex, Cursor, Cline, OpenCode: strip away the terminal UI or the editor pane and each of them serializes a conversation into JSON, POSTs it to a base URL with an API key in a header, and streams the answer back. The client owns the prompt assembly, the tool loop and the rendering; the model sits on the far side of that socket.

Debugging Is Invariant Discovery: What One Kafka Session Taught Us About AI Agents

Tue, 08 Sep 2026 14:14:36 +0000

Every Java team has a sentence like this one somewhere:

Any operation on a TID must go through the single asynchronous path and must not let TMS and EMV drift apart.

It is the kind of rule that is obvious once said and invisible until then. A terminal ID (TID) in a payments system, an asynchronous pipeline that updates two downstream systems, and the requirement that the two never disagree. Nothing in the type system enforces a rule like that, and no single class owns it. Rules of this kind live between consumers, services and stores.

“Commit created” — but it isn’t: why the agent’s word needs external verification, and what the mistake costs

Tue, 08 Sep 2026 08:54:03 +0000

A detailed report reads like proof. The agent lists the files it touched, names the branch, quotes a commit hash, adds a test count and closes with “done”. The work has a shape. The shape looks right. You have six more tickets.

The report is model output. The commit, if it exists, lives in .git. The file, if it exists, lives on disk. The build result lives in a process exit code and a run log. The chat window is none of those places, and it can only vouch for itself.

Did Your AI Agent Run the Debugger? One JVM Bug, Two Agent Runs

Thu, 03 Sep 2026 13:48:32 +0000

Java developers reach for the debugger without thinking about it. Set a breakpoint, run the failing test, look at the variables, then decide what to change.

Most AI coding agents I have watched skip that step. They read the stack trace, read the source, and propose a patch. The reasoning can look excellent and still be a guess about a program nobody ran.

I wanted to know what that guess costs. So I took one real JVM bug, one model, one prompt, and ran it twice with two debugging workflows. Here is what the two transcripts show.