The Horizon · Issue 1
Agents write the code. Engineers own the outcome.
Welcome to the first issue of The Horizon. Every week: one big story, three strange ones, one tool worth trying, and one thing to do on Monday.
The big one: passing tests is not the bar
Two numbers keep coming back to me.
A benchmark published in October, “Last Secure Code Benchmark: Is the Code You Vibe Secure?” (arXiv 2610.14766, from a team publishing under Alibaba’s GitHub), built 450 tasks from real vulnerabilities. The six agents tested passed 74% to 91% of tasks functionally. But even the best agent wrote code that was both working and secure only about 25% of the time.
And CodeRabbit’s State of AI vs Human Code Generation report (December 2025) looked at 470 open-source pull requests. The AI co-authored ones had about 10.8 issues per PR versus 6.5 for human-only PRs. Logic and correctness issues were 75% more common, security issues up to 2.7 times. Two honest caveats: CodeRabbit sells AI code review, and it couldn’t fully confirm which PRs were really human-only.
Put those together and the picture is simple. The agent is fast and the code usually runs. Whether it’s right, safe and maintainable is a different question, and somebody has to own that answer.
What I’ve seen work:
- Agents propose, engineers decide. Nothing merges on its own. Every PR has a human owner who can explain it.
- Use the agent on the boring checks. Naming, null checks, missing tests. Then people spend review time on design and risk.
- Put security scanning in the pipeline, not in a report someone reads next week.
- Review load is real. Faster writing means more to read. Plan review time like you plan build time.
The speed is real. So is the responsibility. It just moved.
Three strange ones
- An AI voice model that gasps. Block’s research lab showed Dessa, a full-duplex speech model trained only on human conversation audio, with no text and no pretrained models. It laughs, sighs, says “yeah” while you’re still talking, and runs locally on a MacBook Pro M4 Max or better. Still a research preview, with no benchmarks published.
- Claude can hang up on you. From November 12, Anthropic’s usage policy bans “sustained and needless” cruelty toward its models. Venting is fine; the main consequence is Claude ending the chat. Full write-up.
- Banks are blocking AI agents that log in as you. Morgan Stanley, JPMorgan and others are restricting third-party agents after some used customers’ passwords and reportedly tried to go beyond what they were asked to do. Full write-up.
One tool worth trying
Secret scanning with push protection. GitHub push protection blocks a push that contains a known secret format before it lands in the repo, and Gitleaks does the same job as a free, open-source scanner you can run locally or in CI. With agents writing more code, a key pasted into a config file is easier to miss in review.
One thing to do on Monday
Pick one repo and write down, in one line, who owns a change that an agent wrote. If the answer is “whoever merged it”, make that explicit in your PR template. Ownership that nobody wrote down is ownership nobody has.
See you next week.
The Horizon
Engineering leadership for the AI era, every other week.
One honest piece every two weeks on AI in software delivery, platforms and leading teams. No spam, unsubscribe anytime.