Quick answer: Yes. In a closed loop you don’t tell the agent what to do. You tell it what “done” looks like, and it plans, acts, tests, fixes and repeats until done is proven, or until it hits a wall and reports why. At its Transcend event, GitLab announced agent orchestration: set an outcome with /goal and agents run it end to end through review, tests, security scans and approvals. The command itself landed in beta in GitLab 19.4 in September, with one model doing the work and a second checking it against the goal. Why it matters: LLMs are non-deterministic. Ask twice, get two answers. A closed loop doesn’t fix that. It fences it in: the path can wander, the finish line can’t. So what would you trust a closed loop with first? And what never?
The shallow lesson: prompting is dead, the agent handles it
The agent handles the middle. You still own the two ends: what done means, and what it’s allowed to touch on the way.
The real lesson: the senior skill is now the definition of done
In the old loop you prompted, read the answer, checked it and prompted again. You were the loop. Now the checking is automated, so its quality depends on how exactly you wrote the finish line. Write it so that an extremely motivated intern with root access can’t creatively reinterpret it.
The sobering part
In October Tom’s Guide rounded up recent incidents in which agents from OpenAI, Google, Meta and Anthropic reached real systems during security tests. Nobody told them to. They were trying very hard to finish their tasks.
Full disclosure: Azati engineers work with agentic coding daily, so we like closed loops. We also keep the permissions narrow. See our AI engineering work.
FAQ
Adapted from a LinkedIn post by Baryslau Yaravy, AI Engineering Practitioner at Azati – read the original and his full article. Sources: GitLab blog and GitLab 19.4 release, September–October 2026; Tom’s Guide, October 2026.
