GitLab Orbit hits GA: agents need 4.5x fewer tokens

GitLab has moved Orbit from beta to general availability. It is a context graph for the entire development lifecycle: code, merge requests, work items, pipelines, vulnerabilities and owners are linked in a single model that updates on every push, merge and deploy. According to GitLab's internal tests, agents connected to Orbit make up to 45x fewer retries and use 4.5x fewer tokens on the same tasks with the same model.

The company's thesis is simple: the bottleneck for agents today is not the model but context. In a large monorepo with tens of thousands of files, an agent keeps going back to the backend for missing data. Every round trip costs time and tokens, and at some point the agent gives up. Worse is a review that looks at a single diff and declares it "safe to merge" without knowing anything about the pipeline or vulnerabilities. GitLab calls this artificial confidence.

One graph query instead of crawling files

When Claude Code is connected to Orbit via MCP, it does not page through files; it queries the graph. The very first response contains the actual callers, recent changes and owners, so follow-up prompts refine the result instead of correcting it.

The figures are backed by a case study from the British aggregator Compare the Market. The team compared four approaches to context for an internal AI reviewer on 79 real merge requests. The Orbit-based version matched roughly 70% of expected inline comments. RAG delivered about 58%, which is worse than having no additional context at all. Ryan Harvey, head of AI Engineering, noted better comment placement and more accurate descriptions of changes.

Token economics: don't pay for repetition

GitLab cites an analysis by a research firm, according to which the cost of an "undisciplined" agent session comes mainly from re-sending context that did not change between turns. This is a property of how the agent is configured, not of the model. Orbit returns a bounded answer from the graph instead of reloading everything that might matter: each turn carries less context, and fewer turns are spent on work that will never reach production.

The graph is shared, so costs do not multiply with the number of agents. Duo agents, Claude Code, Codex and homegrown tools all query the same graph.

Questions that previously had no answer

Beta partners ran queries like these against Orbit:

• Which production services are affected by a vulnerability. One query returns the services with the component, the merge requests that introduced it, the environments and the responsible teams. One partner answered a security coverage question across 34 million entities in seven minutes; via the API, it would have taken about half a day.
• What will break when something changes. Before a push, a developer or agent sees the affected services, jobs and environments across the entire project group.
• Why a CI job fails, and where else. A financial company found the cause of a build failure: the job failed 50% of the time, and no other tool had connected that pattern.
• What depends on a component being migrated. One partner completed an upgrade across several repositories by querying for projects on the old version, with no spreadsheets or repository-by-repository audits.

Setup, scale and pricing

Compatibility:

• GitLab Duo Agent Platform agents work with Orbit natively;
• Claude Code, Codex, Cursor, OpenCode, Gemini CLI, Antigravity and other MCP agents connect via MCP or CLI with the commands `glab auth login` and `glab orbit setup`;
• custom agents work through REST;
• engineers use the same graph through Data Explorer.

GitLab claims that at its scale, it indexes more than 1.5 million projects and 8 TB of graph data in under 2 hours. Indexing runs as a separate service, so the load from agents does not slow down GitLab itself. Permissions are inherited: nested groups, roles and security domains are mirrored and do not need to be rebuilt.

The pricing model is query packs. A pack costs $30 in GitLab Credits per user per month and adds 1 000 queries to the organization's shared pool; seats are not assigned to individual humans or agents. Beyond the pool, you get 33 queries per credit, about $0.03 per query. Promotional pricing for early customers is in effect this quarter.

Orbit is available on the Premium and Ultimate tiers: GitLab.com first, followed by Self-Managed and GitLab Dedicated within several weeks. A 30-day free trial is available through an account manager.

In practice: if the agents in your infrastructure are stuck on repeated requests and expensive long contexts, it makes sense to measure the number of turns and tokens on your own task before and after connecting Orbit. GitLab's figures come from internal tests, so you should run the evaluation on your own monorepo and factor in $0.03 per query beyond the pool.