Agents Broke Git Hosting. The Rebuild Has Started.

The volume broke the tool

Coding agents like Claude Code and Codex went from experiment to everyday tool in less than two years. The volume tells the story: in May 2026, Anthropic merged about 2,500 code changes a day into its monorepo; by September it was merging more than 30,000. Incident trackers counted 48 major GitHub outages in the year ending April 2026, about one a week. That volume is the core of the argument: tools built for human-scale traffic cannot absorb agent-scale traffic without a redesign. m:…5799

Entire is the response from Thomas Dohmke, who left GitHub to start it. With its latest release, the company says, users no longer need GitHub at all: they can host, review, merge, and deploy code on Entire, a Git-compatible forge built for agent-scale traffic. Its hosting runs on a decentralized Git network with repositories copied across the U.S., EU, Australia, and India. m:…5799

Entire is not building its own continuous integration system; instead, it integrates with Buildkite, Depot, and CircleCI. m:…5799

The design details are aimed squarely at agents. Entire's Trails feature replaces traditional pull requests by keeping agent intent and outcome context alongside code changes. Entire Graph maps how a codebase fits together, and Entire Brain links that map to the reasoning behind each change. On ForgeMark, Entire's open-source benchmark, the company says a single repository handled about 570,000 clones an hour and 586 pushes a second. m:…5799

Chaos first, then governance

The chaos is normal, another essay argues. One of the best signs a technology will change the world is whether it makes a chaotic entrance, and PCs, spreadsheets, and email all arrived the same way: employees brought them to work, allowed or not, because they helped get things done. m:…0662

Today, most agents live somewhere chaotic, the essay argues: they run in chat windows, personal accounts, and one-off scripts, holding whatever credentials and data someone handed them. Its prescription is not to suppress that, but to build habitats for company communication and work, with the permissions and access controls to rein in the chaos just enough. Code, it argues, is the ultimate aligning artifact because it literally is the business logic of the organization. m:…0662

It cites a Replit example in which one employee at Sumitomo Mitsui Finance and Leasing took over a development process that was supposed to take six people six months, and did it alone in one week. The measure of a company's alignment, it argues, will soon be the richness of its agent habitat. m:…0662

Agents that use computers directly

OpenAI's computer-using models add a second front. Computer use has reached a place where it is faster than the average human at navigating software, according to OpenAI, and functions as a universal interface: a way for models to operate legacy software that has no dedicated integration tools. m:…7735

The constraint emphasized is control, not capability. Enterprise adoption requires deterministic controls such as sandboxing and human-in-the-loop approval, and OpenAI pointed to its Astra models as more reliable, at 99% versus 97%, and more resistant to prompt injection than prior models. The existence of computer-using agents will pressure companies to make their software more agent-compatible, such as through WebMCP standards. m:…7735