Cursor has given its cloud agents a place to work. The AI coding company rolled out development environments on 13 May that let a single configuration hold all the repositories an agent needs, reusable across sessions. The unglamorous part is the point. An agent is only as useful as the environment it runs in.

What shipped

Three changes stand out. Cloud agents now support multi-repo environments, so one setup can span the several codebases a real task touches. Docker builds got a layer-caching upgrade: change the Dockerfile and only the altered layers rebuild, which the company says makes cache-hit builds 70 per cent faster. And build secrets let an agent reach private package registries during a build without those secrets leaking into the running environment.

The quality-of-life bits

Cursor will now interview you as it sets up an environment, asking questions, flagging missing credentials and checking the configuration works before an agent starts. Each environment keeps its own version history, so a team can review changes and roll back. For teams that would rather not write a Dockerfile, Cursor can inspect the repos, work out the dependencies and generate one to edit.

Why it matters

The gap between a coding agent that demos well and one a team trusts on a live codebase is mostly setup: the right repos, the right secrets, a build that does not take ten minutes. Closing that gap is less exciting than a new model, and more decisive for whether agents get used. Cursor is competing with GitHub, Replit and a field of cloud players on exactly this ground, and the winner will be whoever makes the boring parts disappear.

Three months on, the plumbing turned out to be the asset

The argument here was that environment setup, not model quality, decides whether a coding agent gets used on real work. What happened to Cursor since suggests the industry priced that plumbing considerably higher than anyone was saying in May.

On 8 July, SpaceXAI released Grok 4.5 at US$2 per million input tokens, a coding-focused flagship trained with Cursor data. Six weeks after that, SpaceX closed its acquisition of Anysphere, Cursor's maker, for US$60bn paid entirely in SpaceX stock. It is the largest startup acquisition on record, and Cursor now sits inside a SpaceXAI division.

The sequence is not a straight line. The development environments did not cause the acquisition directly; they let cloud agents run more real tasks against more real repositories. The record of those runs is exactly the kind of data a coding model is trained on. A company that spent its engineering effort making the boring parts work ended up holding something a rocket company paid US$60bn in stock to own.

The competitive ground moved

The contest named in May was Cursor against GitHub, Replit and a field of cloud players, with the win going to whoever made the boring parts disappear. The boring parts are now being commoditised, and the contest has moved somewhere else.

Warp shipped Factories on 20 August, running coding agents as a pipeline rather than a prompt: triage, spec, implementation, review and verification as separate stages, with a different model permitted at each one. Model neutrality is the explicit selling point, and Warp made it days after the Cursor deal put Cursor's own neutrality in doubt. The question about Cursor is no longer whether its environments work, but whether an editor owned by a model company will keep routing to rival frontier models.

GitHub answered on its own terms. Copilot Workspace reached general availability at Build 2026, and Microsoft confirmed Copilot would move to Project Polaris, its own mixture-of-experts coding model, replacing GPT-4 Turbo as the default engine across all subscribers by August. Cognition raised US$1bn at a US$26bn valuation with Devin on a US$492 million annual run-rate. Every player named in May is still standing, but now each of them owns, or is buying, the model underneath.

The bottleneck moved to review

Setup was the constraint this piece identified, and closing it exposed the next one.

Warp's own framing is that a staged pipeline relocates the bottleneck to review capacity. The evidence on how well that review works is not encouraging. A permission-approval study logged 409,000 decisions across more than 40,000 runs and found reviewers missed 33.7 per cent of malicious commands while blocking genuinely safe ones at rates as high as 59 per cent. Anthropic's own figures put the human catch rate at 13.6 per cent against a classifier's 89 per cent, and from 14 August the approval prompt in Claude Code was off by default for Pro, Max and Team accounts.

In May, the gap between a demo-ready agent and one trusted on a live codebase was mostly setup. By August, setup was largely solved. The trust question had relocated to a human review step that measures badly in both directions, prompting several vendors to remove the control rather than improve it.

What a reusable cloud environment costs

A consequence not visible in May is that an environment living in the cloud, reused across sessions, concentrates a team's dependency on its host and on the registries it pulls from during a build. GitHub was degraded for about three hours on 17 August, with archive downloads failing at roughly half, and continuous integration broke in organisations that had never registered they depended on GitHub being available. An agent whose environment rebuilds from a hosted configuration inherits that exposure by design.

This does not argue against reusable environments, but it does mean the reliability question has travelled with the setup question. What used to be one developer's local machine breaking is now a shared dependency breaking for everybody at once.