You open your editor. You describe a task. You wait. You review. You tweak. You commit. Every AI coding tool on the market — Cursor, Claude Code, OpenAI's Codex — starts with the same ritual: a human types a prompt. The AI is fast, sure. But you are still the bottleneck. You're the one who has to context-switch from your issue tracker to your editor, frame the problem in natural language, and babysit the output. Your backlog isn't shrinking because every AI interaction costs you the one resource AI can't manufacture: your attention.

In the first ten days of April 2026, GitHub quietly rebuilt that entire loop — and removed you from the starting position.

On April 1, GitHub rebranded its "Copilot coding agent" to Copilot cloud agent and expanded it beyond pull requests. The agent can now work on branches independently, research a codebase before touching anything, and generate implementation plans before writing a single line. On April 3, GitHub added organization-level runner controls — letting admins set default infrastructure for the agent across every repo and lock it down so individual teams can't override. Same day: cryptographic commit signing, meaning every agent commit now shows a "Verified" badge, which unblocks repos that require signed commits as a security policy. On April 8, the whole thing landed on GitHub Mobile. You can now assign an issue to @copilot from your phone on the subway and get a ready-to-review pull request by the time you reach the office.

Six changelog entries in ten days. That's not a feature launch — that's a platform buildout.

The architectural fork nobody's talking about

Here's why this matters more than any model benchmark. Every other coding agent is prompt-driven: you open Cursor, you type what you want, Cursor does it. You invoke Claude Code in a terminal and describe the task. You queue a job in Codex's cloud dashboard. In every case, a human initiates the interaction.

Copilot's cloud agent is event-driven. You assign @copilot to a GitHub Issue — that's it. The agent reads the issue description, breaks it into a checklist, opens a branch, writes code, pushes iterative commits, runs your automated tests and linters, and opens a pull request requesting human review. No editor session. No terminal. No prompt. The issue itself is the prompt.

The difference isn't the model powering the agent. GitHub routes to the same OpenAI and Anthropic models that competitors use. The difference is workflow position. Issues, pull requests, Actions, code review, and the repository itself are all GitHub surfaces. Copilot doesn't need an integration layer because it already lives inside the system of record. It's not connecting to your workflow — it is your workflow.

The price of removing the human trigger

But let's not pretend this is all upside.

Event-driven agents create a new problem: review fatigue. When a human assigns ten low-priority issues to @copilot on a Monday morning, those PRs land in the review queue whether or not the team has bandwidth to process them. Autonomous output volume can overwhelm review capacity faster than it shrinks the backlog. You traded one bottleneck — prompting — for another: reviewing code you didn't ask for in the moment it arrives.

GitHub seems aware of the pressure. On April 10, it enforced new rate limits for Pro+ users, citing "an increase in patterns of high concurrency and intense usage." It also retired the Opus 4.6 Fast model immediately and paused new free trial signups due to abuse. Translation: people discovered the autonomous coding loop and flooded it.

Meanwhile, Cursor isn't standing still. On April 2, Cursor 3 launched with parallel agent orchestration — multiple agents working on refactoring, testing, and docs simultaneously — plus a dedicated "Agents Window" for managing multi-step projects. It's prompt-driven, yes, but the prompt interface got dramatically more powerful.

What this means for you

If your team already lives on GitHub — issues, PRs, Actions, the whole stack — Copilot's cloud agent is the lowest-friction path to autonomous coding today. No new tool to install. No new interface to learn. Assign an issue, review a PR. The agent works within the governance model your org already enforces: branch protection rules, required reviews, signed commits, runner policies.

If you value model choice, fine-grained control over what the agent does at each step, or you simply don't trust code you didn't explicitly request — prompt-driven tools like Cursor 3 or Claude Code give you more transparency and tighter feedback loops.

Both approaches will coexist. But the direction is clear.

The prompt was the last human bottleneck

For three years, we optimized AI coding around better prompts. Clearer instructions. Richer context windows — the amount of text the AI holds in working memory. Smarter autocomplete. All of it assumed the human presses the first button.

GitHub removed the button. The developer's primary job just shifted — quietly, across six changelog entries in ten days — from "tell the AI what to build" to "review what the AI already built."

Whether that's liberation or a new kind of hell depends entirely on how good your code review process is. And if you've ever worked on a team with a 200-PR backlog... well. You already know the answer.