GitLab Orbit GA Claims 45x Fewer Agent Retries

GitLab has made Orbit, a context graph that gives AI coding agents a live view of a team’s code, merge requests, pipelines, vulnerabilities and owners, generally available. In GitLab’s own testing, it says agents connected to Orbit needed up to 45 times fewer retries and used up to 4.5 times fewer tokens on the same tasks with the same model.

The company announced the release on 29 October in a blog post and on X. Orbit is available now to Premium and Ultimate customers on GitLab.com, with Self-Managed and GitLab Dedicated customers due to get it “over the coming weeks”. It has been in beta since June.

What Orbit does

GitLab’s argument is that coding agents fail less because of the model and more because of what they can see. An agent reading a single diff does not know which merge request last changed the code, which pipeline ran it, which vulnerability it carries or who owns it. So it searches, asks again and burns tokens.

Orbit pulls that information into one queryable graph built from data already held in GitLab, with existing group and role permissions mirrored. GitLab’s own agents on the Duo Agent Platform query it natively. Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Antigravity and other agents that support the Model Context Protocol (MCP) can connect through MCP or the GitLab command-line tool, using glab orbit setup. Custom tools can use a REST interface, and engineers can query the same graph through a Data Explorer.

The pitch fits a wider shift towards agents that open and test changes on their own, such as Replit’s cloud agents and Cursor’s Agent SDK.

The customer test

GitLab’s main outside example comes from Compare the Market, the UK price comparison site. According to GitLab, Compare the Market tested four ways of giving context to an internal AI code reviewer across 79 real merge requests. The Orbit-based reviewer matched about 70% of the inline comments expected, GitLab says, while a retrieval-augmented generation (RAG) setup managed about 58%, which GitLab says was worse than giving the reviewer no added context at all.

“We tested it against RAG and a few other approaches across real merge requests, and the gap was clear,” Ryan Harvey, Compare the Market’s head of AI engineering, said in GitLab’s post.

GitLab also cites unnamed design partners. One, it says, answered a security-coverage question across 34 million entities in seven minutes, a job it says would have taken half a day through the API. A financial services company traced a build-system outage to a CI job that failed 50% of the time, according to GitLab.

Pricing

Orbit is sold as query packs. Each pack costs $30 in GitLab Credits per user per month and adds 1,000 queries to a pool shared across the organisation, by people and agents alike. Usage beyond the pool is billed at 33 queries per GitLab Credit, which GitLab puts at about $0.03 a query. Customers can try it free for 30 days through their account team, and GitLab says promotional pricing is available this quarter.

Vendor benchmarks need independent repeats

This is TSN’s analysis. The headline numbers are GitLab’s own, from its own tests, and GitLab has not published the tasks, the model, the agent set-up or the baseline behind “up to 45x fewer retries”. “Up to” means the best case, not the typical one. The 4.5x token figure is more modest and more plausible, since a single bounded answer from a graph should cost less than an agent rereading files. Without the method, though, neither number can be checked.

The Compare the Market result is more useful because it involves a named customer, a stated sample of 79 merge requests and a clear comparison. It is still one company, one internal reviewer and one metric, reported by the vendor. A gap of about 12 percentage points over RAG is meaningful if it holds, but it needs repeating on other codebases.

The pricing is also worth working through. At $30 for 1,000 queries, a pack works out at $0.03 a query, the same as the overage rate. A busy agent that asks the graph dozens of questions per task could use a pack quickly, so the shared pool matters more than the per-user label suggests.

There is a security angle too. A single graph linking code, vulnerabilities, environments and owners is very useful to defenders, and equally valuable to an attacker who gets in. GitLab itself recently fixed a second 10.0-rated code-execution flaw in 10 days, and Alpine Linux rebuilt its GitLab after an attack. GitLab says Orbit mirrors existing permissions, so the safety of the graph depends on those permissions being right. As Microsoft opens its own agent harness to developers, the race is shifting from which model to use to which platform holds the context, and buyers should ask for reproducible benchmarks before paying for either.