Git works extremely well for human developers. But when AI agents continuously generate reads, writes, checkpoints, tests and merges, the workload changes. GitHub and GitLab are now adapting their infrastructure to that new reality.
Why agents change Git workloads
A human developer rarely creates a commit after every small step. An agent can read a few files, change code, create a checkpoint, run tests and repeat seconds later.
GitHub says agent-scale development is increasing concurrent writes, automated reads and merges. In September 2026, GitHub reports processing 7.38 billion commits, more than five times the volume a year earlier. Monthly Git events reportedly grew from about 218 billion to 473 billion between September 2025 and August 2026.
Writes are harder to distribute than reads
Files and clones can be accelerated with caches and replicas. A write must be durably recorded and then become consistently visible to other agents, CI systems and developers. Concurrent writes therefore become a major scaling challenge.
Merges converge on the same place
A hundred agents can work on a hundred independent branches. When those branches need to reach main, operations converge on the same reference.
GitHub says pull request merges have nearly quadrupled in a year, making merge queues and validation strategies increasingly important.
Every change triggers a cascade
A push can trigger tests, builds, security scans, indexing, notifications and sometimes new agents. GitHub reports that GitHub Actions ran 3.26 billion jobs in September 2026.
With agents making frequent small changes, selecting only the necessary validation becomes increasingly important to keep CI cost and latency under control.
GitLab is reaching the same conclusion from another direction
GitLab is also working on a new generation of source code management designed for agents. One concern is clone cost: an ephemeral agent does not always need an entire repository just to inspect a few files or change a few lines.
GitLab is experimenting with more server-side operations so agents can request precisely the data they need without rebuilding the full context locally.
The underlying shift
Git remains the versioning model. What is beginning to change is how agents interact with Git.
Fewer clones, more specialized interfaces
For humans, a complete distributed copy of a repository remains extremely powerful. For a short-lived agent, a targeted interface can reduce network traffic, startup time and loaded context.
The interaction may gradually move from agent → Git directly to agent → SCM service / API → Git.
Identity and traceability become central
A change can now be produced at a user's request, by an agent, a sub-agent or a CI workflow, using a particular model and permission set.
The question may no longer be only “who made this commit?” but also: which agent, which task, which model, which permissions and which original request?
What this changes for development teams
- More ephemeral branches: creation and cleanup need to stay fast.
- Merge queues: coordination matters when many changes converge on the main branch.
- More selective CI: rerunning every test after every micro-change can become expensive.
- Stronger traceability: a workflow's technical identity may no longer fully explain where a change came from.
What this suggests for TechAtelier
The interesting signal goes beyond Git: tools designed around human activity are starting to be adapted for much faster and more parallel machine activity.
The same effect can reach CI/CD pipelines, APIs, alerting, monitoring, quotas, security policies and audit logs. An architecture that works perfectly for a few human users can behave very differently when agents perform hundreds of operations automatically.
What to remember
Git has not suddenly become obsolete. The usage profile is changing.
GitHub is rebuilding parts of its Git infrastructure for this new workload, while GitLab is developing an SCM backend oriented more toward server-side queries and agent traceability.
When multiple major platforms start changing their architecture around the same phenomenon, this is more than another AI feature: software development itself is beginning to operate at a different scale.
Sources
- GitHub — Building Git infrastructure for agent-scale development
- GitLab — Next-generation source code management for agents
Published October 7, 2026. Figures and features reflect information published by GitHub and GitLab as of that date.