30,000 merges a day: AI hits the software delivery bottleneck

A change is a unit of work that has been reviewed and merged into the shared codebase. Git was designed for human speed: a developer writes code, a colleague reviews it, CI runs tests, and the branch is merged. Thirty thousand daily merges mean about 21 changes per minute around the clock.

Each change targets the same main branch, reruns checks and may conflict with neighbouring changes.

Willison marked public Claude release dates in 2026 on Anthropic’s chart.

He notes that internal models appear before public releases, so connecting the jumps to specific versions remains a hypothesis.

The main conclusion stands: at a company that develops frontier models itself, version control became the bottleneck. The models handled the workload.

A company with twenty developers will not have 30,000 merges a day.

Its bottlenecks are the same, on a smaller scale. ### Merges and environments

If agents triple pull-request volume, the build queue will grow faster than CI capacity.

Changes will wait for hours on a shared test environment. ###

Review: a team lead who reviewed ten pull requests yesterday now receives thirty.

The lead either lets changes through unchecked or slows the entire team. Either way, the benefit of generation disappears. ### Software supply chain

The crates.io security team warned of a targeted campaign against maintainers of popular Rust packages.

Attackers arrange a video call under the pretext of a job or contract, then ask the victim to install a missing audio codec.

They take over the device and account to publish an infected package version. An agent allowed to add dependencies independently can pull it in that same hour unless a filter separates it from the registry.

In AI-native development projects, we put agents through the same delivery process as human-written code.

An agent’s change passes automated tests, static analysis, dependency allowlist checks and a merge queue.

The agent receives a task with acceptance criteria. A person checks the result against those criteria rather than rereading every line of code.

Model requests pass through an LLM & Security Gateway: a single access point for Claude, Gemini, GigaChat, YandexGPT and Qwen.

The gateway logs calls, masks personal data and enforces budget limits.

Changing a model in the gateway takes one configuration line, allowing a team to test Argon at $2 per million tokens on its own tasks within a day.

We expose 1C, Akeneo or Pimcore PIM, and Apache Kafka to agents through MCP tools with explicit permissions for each action. An agent enriching product descriptions, facets and taxonomy writes to PIM only after schema validation.

Our editorial agent follows the same rules when preparing this blog. It interacts with the site through Git, runs source checks before publication and commits only its own files.

Generated lines and pull-request counts measure activity.

TTU measures the result: time from a task being assigned to a working production feature.

Track three other numbers alongside it: review-queue time, build time and the share of changes rolled back.

If AI triples code output but TTU stays unchanged, the company is paying for theatre.