GitLab is extending its development platform in version 19.2 with four so-called agentic functions. In this context, agentic means that a language model not only outputs suggestions in a chat window but independently executes several work steps in the platform – for example reading code, generating a change and submitting it as a merge request for review. Two of the four functions are generally available, two are available to try out as a public beta.

What the four functions address

The focus is on tasks that regularly arise in projects, rarely get priority and are nevertheless security-relevant:

  • Update vulnerable dependencies: Third-party libraries for which a security vulnerability has been reported should automatically be raised to a non-critical version.
  • Check code changes for logic errors: The analysis goes beyond formal checks such as linting and looks for errors in the program logic itself.
  • Automated processes in the terminal: Command-line steps that were previously issued manually can be delegated to an agent.
  • Automated workflows: Recurring processes within the platform are chained together and processed independently.

What is remarkable is less the individual capability than the place: The functions sit in the platform where repository, pipeline, merge requests and security reports come together anyway. This eliminates the detour via additional tools that first need access to the source code – a point that experience shows costs a lot of time in tenders and data protection reviews.

Why dependency updates are the most interesting case

Outdated dependencies are the biggest open construction site in many mid-sized projects. Not because nobody knows about them, but because an update rarely fits into a sprint budget: raise the version, read the changelog, run tests, fix follow-on errors. This very chain is mechanical enough to give it to an agent, and important enough not to leave it lying.

The gain here lies not in pure speed but in regularity. A project in which security updates are continuously rolled out in small steps remains updatable. One in which nothing happens for two years becomes a migration project with its own quote.

Logic checking does not replace review

Restraint is appropriate with the automated checking of code changes. A model can name conspicuous patterns, address edge cases and mark places that a human would have overlooked. Whether a change does the right thing in terms of the domain, however, is decided by requirements that are not in the repository. The sensible use is therefore the upstream pass: The agent clears away the obvious points, the human review concentrates on architecture and domain logic.

The same applies to terminal and workflow automation. Anyone who lets agents execute commands shifts a question from technology to organisation: What rights does this agent have, which environments may it touch, and who confirms a step before it takes effect in production? For this reason, public beta functions belong first in side projects or in clearly delimited repositories, not in the pipeline of the revenue-bearing system.

What teams can do concretely now

  • Check which of the four functions are available at all in your own licensing and operating model, and separate the beta functions from the generally available ones.
  • Select a pilot project with manageable risk and automate only dependency updates there.
  • Clarify beforehand how robust the test coverage is. Automatic updates only unfold their benefit once a pipeline reliably says whether something has broken.
  • Establish that agents submit suggestions via merge requests and do not write directly to protected branches.
  • After a few weeks, measure how many suggestions were adopted without rework – that is the only reliable basis for an expansion.

Assessment

For web and software projects, this shifts less the question of whether AI participates in the development process than the question of where the control points lie. The practical lever lies in maintenance: If security updates and routine checks run reliably in the background, more of the agreed budget remains for functions that are visible to the business. The prerequisite remains a pipeline with meaningful tests and clearly defined approvals – without them, you only automate the speed at which errors enter the system.

Sources