GitLab Orbit 正式推出，整合軟件生命週期資料供 AI 智能體查詢

GitLab Orbit Is Now Generally Available: One Context Graph for Your Whole Codebase, Built for AI Agents

GitLab announced today, 29 October 2026, the general availability of GitLab Orbit — a context graph for the software lifecycle. It connects code, merge requests, work items, pipelines and vulnerabilities into a single, continuously updated view that AI agents and engineers can query, instead of hunting through files one by one.

Nothing here is sponsored, and there are no affiliate links on this page. Every fact below comes from GitLab's announcement of 29 October 2026 (BusinessWire). I have not tested Orbit myself — this is researched analysis, not hands-on testing. Where a number is GitLab's own claim, I say so plainly.

The problem Orbit is trying to solve

If you have ever watched an AI coding agent work in a large codebase, you have seen the failure mode: the agent does not know enough about the project, so it burns iterations re-searching files, asks the wrong questions, and comes back with answers that are inaccurate or incomplete. The context it needs is scattered — some in the code, some in merge requests, some in issue trackers, some in pipeline runs, some in vulnerability reports.

Orbit's answer is to stop making the agent go and find all of that. It builds one queryable view of the whole lifecycle and keeps it updated as things change. One graph, one query, instead of a scavenger hunt across five systems.

That is a genuinely different pitch from "a smarter model." It is not claiming the agent got cleverer — it is claiming the agent finally got the full picture. Whether that picture is as complete and as fresh as promised is the thing real-world use will prove or puncture, and no launch-day announcement can settle it.

What Orbit actually does

One view across the lifecycle. Code, merge requests, work items, pipelines and vulnerabilities, connected in a single context graph that updates continuously. An agent — or an engineer — queries the graph rather than re-discovering relationships file by file.

Built for agents first, engineers second. The graph is designed to be queried by AI agents, but the same unified view is useful to a human trying to understand why a change was made, what it broke, and what is still open. If you have ever traced a bug through a merge request, a pipeline failure and a vulnerability report in three different tabs, you know why one view matters.

Native to GitLab Duo, open to other agents. Orbit works natively with the GitLab Duo Agent Platform, and third-party agents are supported too — Claude Code, Codex, Cursor and OpenCode — via MCP or the GitLab CLI. That last part matters: GitLab is not locking the graph to its own agent. If your team already runs Claude Code or Cursor, Orbit is meant to feed those tools, not replace them.

Already battle-tested at scale — by GitLab's account. Since its introduction in public beta at GitLab Transcend on 10 June 2026, GitLab says more than 3,500 organisations have used Orbit, and agents have run more than 280,000 queries through it. That is a real beta footprint, not a handful of design partners — though it is still GitLab reporting its own numbers, and independent verification is not available yet.

The headline numbers — labelled honestly as GitLab's claims

GitLab's own testing claims up to 45x fewer hallucinations, up to 4.5x fewer tokens, and fewer retries. Read those words carefully: "up to", and "GitLab's own testing". These are the company's claims from its own evaluation, not independent benchmarks, and this article does not present them as verified results.

That said, the direction of the claims makes sense. If an agent stops re-searching files and stops guessing at missing context, you would expect fewer hallucinations, fewer wasted tokens and fewer retries. The mechanism is plausible. The magnitude — 45x, 4.5x — is what independent testing needs to confirm, and until it does, treat those numbers as a claim to check, not a fact to quote.

Availability and pricing — what is confirmed, and what is not

Who gets it: Orbit is generally available for Premium and Ultimate customers.

Where and when: starting on GitLab.com today, 29 October 2026, and reaching Self-Managed and GitLab Dedicated over the coming weeks. If you self-host GitLab, you are waiting — the announcement does not give a firm date for your deployment, only "over the coming weeks".

How it is sold: in query packs, sized by GitLab account teams. There is no public per-query or per-pack price in the announcement, and I am not inventing one. If a blog post quotes you a price for Orbit today, it is guessing.

Trying it free: a free 30-day trial is available via your GitLab account team — not, as far as the announcement states, a self-serve signup. Talk to your account team if you want to evaluate it.

Who this is actually for

Teams already deep in GitLab: if your code, issues, pipelines and vulnerabilities already live in GitLab, Orbit connects what you already have. The value proposition is weakest for teams whose lifecycle is spread across many vendors.
Teams running AI agents in large codebases: small repos do not have a context problem worth solving this way. The pain Orbit targets — agents lost in a big codebase — grows with codebase size and team size.
Teams already using Claude Code, Codex, Cursor or OpenCode: MCP and CLI support means you do not have to switch agents to try the graph. That lowers the trial barrier considerably.
Probably not yet for: Self-Managed and Dedicated customers (wait for the rollout), teams outside the GitLab ecosystem, and anyone who needs independent benchmark proof before committing. The 45x and 4.5x figures are GitLab's, not a third party's.

What we don't know yet — stated plainly

Independent verification of the 45x / 4.5x claims: not available at publication. GitLab's own testing is the only source so far.
Pricing: no public price for query packs was announced. Do not trust a quoted price that does not come from GitLab.
Self-Managed / Dedicated dates: "over the coming weeks" is the only timing given — no firm date.
Real-world performance outside GitLab's beta: 3,500 organisations and 280,000+ queries is GitLab's beta account. How the graph holds up under general-availability load, and how much it helps agents on codebases unlike the beta cohort's, will emerge from independent use.
How it compares: this is not the same story as GitHub's Copilot announcements from Universe week. Orbit is a context layer, not a model or an agent — judge it on whether agents get better answers with it, not on model benchmarks.

None of that takes away from the structural point: AI agents have been starved of context, and a continuously updated graph of the whole lifecycle is a serious attempt to fix the starvation rather than just building a hungrier agent. If the claims hold up even approximately in independent use, this is one of the more consequential developer-infrastructure releases of the year. But the honest version today is that the mechanism is sound, the beta footprint is large, and the proof is still GitLab's own — independent testing comes next, and I will update this article when it lands.

Sources: GitLab announcement via BusinessWire, "GitLab Orbit generally available", 29 October 2026 (announced 13:00 UTC). All figures attributed to GitLab as stated in the announcement.

Full version with FAQ: https://toolscout-e2i.pages.dev/posts/gitlab-orbit-generally-available-ai-agent-context-graph.html