According to GitLab, agents are predominantly performing unnecessary work. Due to an abundance of context, organizations use 4.5 times more tokens than necessary, the company states. The newly released GitLab Orbit, equipped with a context graph, is expected to reduce the number of AI searches for the correct context by up to 45 times.
This context graph is said to encompass all of an organization’s code, regardless of its place in the software lifecycle. Every relationship would be traceable to provide the correct context for the information. According to GitLab, agents are typically bound to the diff they find, but they need pushes, merges, and deployments just as much.
In a large monorepo, an agent continuously queries the backend, GitLab explains. Each roundtrip costs time and tokens, until the agent eventually gives up. “The model isn’t the limitation. Context is,” the company states. But a context graph alone isn’t enough: an entirely new system would only introduce more complexity. That’s why GitLab’s innovation builds on existing information.
One graph with many agents
Agents on the Duo Agent Platform use Orbit without requiring any additional configuration. Claude Code, Codex, Cursor, Gemini CLI, and other MCP-compatible agents connect via MCP or the GitLab CLI. Orbit mirrors existing groups, roles, and access permissions for use within the context graph. According to GitLab, it’s important that agents and engineers see the same graph to ensure consistent results.
Orbit is offered in so-called query packages, which collectively form a single shared pool for the entire organization. A package costs $30 in GitLab Credits per user per month and provides 1,000 queries. Additional usage costs approximately $0.03 per query.
Orbit is available to Premium and Ultimate customers, with initial access on GitLab.com. Self-Managed and GitLab Dedicated will follow in the coming weeks. Customers can try it for free for 30 days.
See also: GitLab frees up funds for agentic AI investments through layoffs