91% of organizations now run two or more AI coding tools, and 54% run three or more, but only 21% report productivity gains across the full software lifecycle. GitLab’s own research across 1,500 developers and tech leaders confirms what engineers feel: more agents, more code, more chaos. Some customer codebases grow five times in a single year. Unmanaged agent concurrency turns speed into churn.
That’s the backdrop for everything GitLab announced at its Transcend event. The headline is a rebuilt Git engine — next-generation source code management, now in private beta — designed from the ground up for agent-scale concurrency, not human-scale clones. The old model breaks when each team member runs hundreds of agents: clone tax on every read, concurrency collapse under thousands of sessions, and no agent isolation. The new backend stays compatible with the Git protocol but swaps the motor. Early internal results show up to 2x fewer tokens consumed, up to 50x faster wall clock time, and up to 1,000x less network traffic. That’s the difference between a system that tolerates agents and one that expects them.
Context Graph That Cuts Hallucinations 45x
Agents see code but not the lifecycle around it — dependencies, pipelines, deployments, incident signals. That blind spot wastes tokens and produces code that looks correct but gets reverted. GitLab Orbit, now in public beta, maps the full software lifecycle into a single live graph. Agents query that graph instead of scraping fragmented context. In internal tests, Orbit-grounded agents delivered up to 11x faster response, were up to 4.5x more cost effective, and hallucinated up to 45x less. Customer Compare the Market A/B tested Orbit against a RAG-based baseline and a no-context control, using identical prompts and models across 79 real merge requests. Graph-grounded agents placed inline review comments in the correct location 70% of the time (0.696 accuracy) versus 58% (0.577) for RAG. That’s a measurable 12-point gain on a single concrete task — comment placement — from a better retrieval architecture.
Governance That Tracks Every Agent Action
When agents push code at scale, the question isn’t “did we scan?” but “which agent did what, under which policy, and can we prove it?” GitLab’s agents for security, available now in GitLab Ultimate, scan, triage, and resolve vulnerabilities as fast as agents create them. Governance for agents, in private beta, wraps identity, policy, audit, and approval around every agent action, with real-time visibility into inputs, reasoning, tool calls, and anomalies. The audit trail captures what changed, which policy allowed it, and who signed off. This is the difference between shipping fast and shipping safely.
GitLab Duo Agent Platform, generally available since January, orchestrates the full loop: picking up issues, opening merge requests, reviewing against team-specific rules, and fixing failing pipelines without pulling engineers out of flow. Weekly active users have grown 10x since GA. GitLab Duo Code Review placed in the top five of the Martian Offline Benchmark for AI code review tools. Separately, GitLab Flex lets teams reshape seats, AI usage, and capabilities month to month under a single annual commitment — no new procurement cycle when adoption shifts unpredictably.
The engine, the context, the governance, and the buying model all point in one direction: turning agentic coding from a productivity gamble into a manageable system. GitLab’s bet is that the platform already sitting in the path of every commit and pipeline is the natural place to anchor that system. The numbers from Orbit and the new SCM suggest they're not just repackaging — they're rebuilding for the concurrency regime that’s already here.
Source: GitLab: Built for the agentic engineering era
Domain: about.gitlab.com
Comments load interactively on the live page.