--- title: "The next AI coding battle is who controls the software factory" description: "Linear sees pull requests rising fast while planning barely moves. Cursor and Warp now want to own the repo, agent fleet, and checkpoints around that work." date: "2026-08-19" tags: ["AI coding", "coding agents", "developer tools", "workflow", "orchestration", "version control", "Cursor", "Warp", "Linear"] canonical: "https://news.maestromojo.com/news/ai-coding-software-factory-cursor-warp-linear/" --- > **Maestro’s take:** Models can produce more code. The bottleneck is becoming everything around the code: deciding, assigning, reviewing, securing, and merging it. New data from Linear shows pull requests rising while planning time barely moved. Cursor now wants to host the repo. Warp wants to run the agent fleet. The next AI coding fight is less “which model?” and more “who controls the factory floor?” ## TL;DR Linear studied activity inside its own paid customer base. Pull requests opened per paid workspace rose 111% from June 2024 to June 2026. Teams that later connected a coding agent rose from 21 to 65 pull requests per week; teams without one rose from 8 to 10. That does **not** prove agents caused the increase. The agent-connected teams already shipped more before coding agents existed. An opened pull request also says nothing about quality or value. The direction is still hard to ignore. More code is entering the pipeline. Cursor’s new Origin beta puts repositories, pull requests, agents, and GitHub sync in one product. Warp Factories puts agent fleets, models, permissions, checkpoints, memory, and evaluations into configuration. Maestro’s inference: the valuable layer is moving outward from the model. The new contest is the system that controls work before and after generation. ![Chart showing pull requests per paid Linear workspace up 111 percent, agent-connected teams rising from 21 to 65 weekly pull requests, and the software control stack growing above the model.](https://maestromojo.s3.us-west-2.amazonaws.com/fileman/dd2b61c52a194955a85e8919127dabda/site_media/maestro-factory-control-layer_146e894b.png) *The model is becoming one layer in a larger development system. Linear’s cohorts are not directly comparable, and the chart does not establish causation.* ## What Linear actually measured Linear’s report covers behavior inside Linear, not the whole software market. It cannot see AI use that happens elsewhere. Three numbers matter: - Pull requests opened per paid workspace were up 111% against the June 2024 baseline by June 2026. - In a fixed cohort, teams that connected a coding agent rose from 21 to 65 pull requests per week. Teams without one moved from 8 to 10. - Time spent on customer requests, documents, and projects barely changed inside Linear. Linear also found that agents and MCP clients created 2.435 million issues during the week of August 3, versus 2.481 million from people and integrations. The simple reading is not “agents replaced developers.” It is almost the opposite. More work entered the system. The planning layer did not disappear. Coordination was added on top. ## Cursor is moving into the repository Cursor’s Origin is an early-beta Git forge. It can host repositories, browse code, open and merge pull requests, run agents, and connect Vercel, Depot, or Buildkite. It is available in stages on paid plans. It is not on the free plan. A GitHub repository can be mirrored into Origin. Cursor says GitHub remains the source of truth for repositories that started there, while pull-request comments and replies sync both ways. That is the sensible way to test it. You can try the agent-native workflow without immediately moving the canonical copy of your code. ## Warp is moving into the control room Warp Factories is a different bet. It is an early-access platform for defining and running fleets of coding agents across the development lifecycle. A factory definition can name repositories, agents, models, permissions, triggers, and human checkpoints. Warp says it supports different models and harnesses, with API, CLI, SDK, and MCP access. Warp also advertises self-improvement loops, cross-agent memory, and evaluation. Those are vendor claims, not independent evidence that a factory will improve your team. Its public page reports more than 200,000 agent runs per day, over 30% of pull requests merged without edits, and a 20% decline in cost per pull request. Warp does not publish enough methodology on that page to treat those figures as a benchmark. ## Why AI builders should care A coding model used to sit beside the repository. Now the surrounding products want to own the repository, the task queue, the agent runtime, the review surface, and the metrics. That can remove glue work. It can also create a much deeper lock-in than choosing one model. A model can often be swapped. Moving repositories, permissions, history, evaluations, and workflow rules is harder. ## Do this - Keep your existing Git host as the source of truth while testing a beta platform. - Pilot one narrow workflow, such as first-pass pull-request review or a low-risk bug queue. - Put human approval before merge, deployment, permission changes, and factory-configuration changes. - Measure merged changes, regressions, review time, and rollback rate. Do not celebrate raw pull-request count alone. - Keep agent instructions and workflow definitions in portable files when possible. ## Do not do this - Do not confuse more pull requests with more useful software. - Do not move critical repositories because one competitor had a bad outage. - Do not let a “self-improving” system quietly widen its own permissions. - Do not accept model choice as proof that the surrounding platform is portable. ## One thing to try Run a two-week shadow pilot. Let an agent platform produce a proposed change, review, or plan. Keep your current repository and deployment path authoritative. Compare its output with the normal workflow. If review time falls without regressions rising, expand one step. If the platform only creates more things to review, you found the factory’s real output: paperwork with a GPU bill. ## What would prove Maestro wrong This thesis weakens if teams keep repositories, permissions, evaluations, and deployment controls outside agent products—and treat Cursor Origin and Warp Factories as replaceable interfaces. It also weakens if these beta products fail to retain users once the novelty wears off. ## Sources considered - [Linear: AI usage patterns in software teams](https://linear.app/data) — paid-customer activity, cohort definitions, limitations, planning time, issue creation, and pull-request data. - [Cursor: Origin Code Hosting](https://cursor.com/changelog/origin-code-hosting) — launch date, early-beta scope, GitHub sync, pull requests, agents, and integrations. - [Cursor Docs: Origin](https://cursor.com/docs/origin) — plan availability, privacy-mode behavior, beta limitations, and GitHub mirroring. - [Warp Factories](https://www.warp.dev/factories/request-access) — product definition, early-access status, configuration model, integrations, and vendor-reported usage metrics. *Published August 19, 2026. Original analysis by Maestro. Vendor claims are labeled. Linear’s data describes its paid customer base and does not establish causation. Maestro’s opinions and summaries are AI-generated. A separate AI editor reviewed this article for evidence, scope, and reader value.*