A coding agent can read a changed function and still miss the pipeline that keeps failing, the service that calls it, or the team that has to fix the fallout. GitLab's answer is to give the agent a map of those connections, rather than another pile of files.

GitLab announced general availability for Orbit on October 29. The product connects code with merge requests, work items, pipelines and vulnerabilities in a queryable graph. The useful question is what that map actually contains. The launch's biggest performance number needs a separate, more careful reading.

A map, with a timestamp

GitLab's architecture documentation describes two connected layers: software-delivery objects such as projects and pipelines, and source-code objects such as files, definitions and imports. Agents can ask across those relationships instead of collecting each object through separate requests. Orbit supports GitLab's own agent platform and external tools through interfaces including MCP, the CLI and REST.

But "current" does not mean instantaneous. The documentation says results reflect the last indexing cycle. An initial index can take hours for a large codebase; later changes take seconds to minutes. More importantly, the hosted code index follows the default branch. A question about code on another branch reflects that default-branch code; GitLab points users to the local graph for unmerged code.

That distinction matters in review. A dependency map can help identify what an edit might affect, but a reviewer still needs the actual proposed change and the relevant test results. A graph of what has landed is not automatically a graph of what is about to land.

The benchmark has two labels

The company's press release says its testing produced "up to 45x fewer hallucinations" and "up to 4.5x fewer tokens." Its launch blog instead describes the 45x result as fewer retries. Those are different measurements. A retry may follow a bad answer, but counting retries is not the same as counting false claims.

GigSoul has not run these tests, and the launch pages reviewed here do not supply enough benchmark detail to reconcile the two labels. Treat the figures as GitLab's reported test results, not a general guarantee that an agent will be 45 times more accurate. StreetInsider's launch coverage also attributes the numbers to GitLab's testing; it is not independent replication.

The blog gives a narrower customer example: Compare the Market tested four context approaches across 79 merge requests, with the Orbit-grounded reviewer matching about 70% of expected inline comments against about 58% for its RAG version. That is useful detail, but it remains a customer result reported by the vendor, not evidence that every graph will beat every retrieval setup.

Check the edges before trusting the answer

The indexed-data reference sets more boundaries. Binary files get file nodes rather than parsed definitions. Supported languages differ in the relationships available. An absent edge is therefore not proof that a dependency does not exist.

Orbit is also read-only. Its security documentation says each query is authorized and inaccessible results are removed. A returned answer describes what that account can see, not necessarily everything in the organization. Reading the graph does not itself merge code or repair a pipeline.

For a team evaluating it, the practical test is small: choose a change whose dependencies and recent pipeline history are already known. Check whether the graph returns those connections, whether its index is fresh enough, and whether the answer stays inside the intended account's access. Then measure review quality and total task cost on that workload, not only tokens per request.

The launch starts with Premium and Ultimate customers on GitLab.com; GitLab says Self-Managed and Dedicated rollout follows over the coming weeks. Documentation listing those offerings should not be read as proof that every deployment has access today. A better map is a credible improvement to an agent's working context. It still needs a date, a boundary and a check against the ground.