AI Coding Agents Didn't Get Adopted. They Got Defaulted.

Cursor v0.46 made AI coding agents the default interaction mode in February 2025. And GitHub switched its Copilot Coding Agent on for every paid subscriber in September 2025.

That is the entire adoption story.

Nobody chose agentic coding through procurement or a benchmark bake-off; vendors changed what happens when you press Enter. And everyone else inherited the result. A January 2026 Faros.ai guide names Cursor, Claude Code, Codex, GitHub Copilot. And Cline as the front-runners. And calls Cursor "the most broadly adopted AI coding tool among individual developers and small teams according to Reddit." If you run a small shop, the question is no longer whether to try an AI coding agent.

It's whether you've noticed your editor and your repo host already ship with one switched on.

Cursor Made the Agent the Front Door

The pivot happened in Cursor v0.46, released February 2025. Chat, Composer, and Agent got unified into a single interface, and Agent became the primary mode. StronglyTyped confirms the same shift: from v0.46 onwards, rolling out late February 2025, Agent mode is the default.

Cursor still technically offers three modes. Verdent's guide describes them as Ask (conversational chat, no edits), Agent (autonomous multi-file work, the default). And Manual (inline edits with explicit control). But the mode you land in when you open the product is the autonomous one. An April 2026 dev.to tutorial puts it plainly: "By default, the Chat will be in Agent mode. So the agent will directly go ahead and write code as you write prompts."

Then Cursor 2.0 dropped the pretense. TechSifted's breakdown of the release says Cursor "redesigned the entire editor around agents rather than files," and calls that a philosophical shift, not just a UI update. You describe the task, agents decide which files to touch. Multiple agents can run in parallel using git worktrees. And background agents run in isolated VMs with internet access and open pull requests when they finish.

Here's my read: the framing flipped. Cursor used to be an editor with an agent inside it. Now it's an orchestrator that happens to render files. A June 2026 dev.to mapping of desktop and cloud coding tools labels Cursor "the most-used AI IDE," a VS Code fork with agent and tab completion, "default for engineers who want a familiar IDE that just got smarter." That last phrase is doing quiet work. The IDE looks familiar. The behavior underneath it no longer is.

GitHub Turned the Agent On for Every Paid Repo

GitHub took a different route to the same destination.

The Copilot Coding Agent became generally available to all paid Copilot subscribers in September 2025. And it is not an IDE plugin. It lives on GitHub itself, in the flow of issues, pull requests, and GitHub Actions. Assign an issue to `@github-copilot` in the sidebar.

And the agent applies changes in an ephemeral environment and updates a draft pull request.

The official docs say the cloud agent is available for all paid Copilot plans, inside Visual Studio Code, JetBrains IDEs, Eclipse.

And Visual Studio, with a "Delegate this task to the GitHub Copilot cloud agent" button sitting in Copilot Chat.

Coverage extends to all repositories stored on GitHub, except repos owned by managed user accounts and repos where it has been explicitly disabled.

Read that exception clause again, because it inverts the old model. The agent is present unless you turn it off. Absence now requires an act. GitHub kept pushing in May 2026, bringing the Copilot CLI agent into JetBrains IDEs in public preview with workspace isolation, plus a unified sessions view showing live status for running and queued agent sessions. You can define custom agents in a `.github/agents` folder per workspace, or globally via an `.agent.md` file under `~/.copilot/agents`. Agent config is becoming as ordinary as a `.gitignore` entry.

For small teams this is the part to sit with.

Most two-person shops have not written a policy on whether a cloud agent may open PRs in client repos.

The default answer, as of today, is yes unless someone says otherwise. That's a decision your shop made without anyone in your shop making it.

The Model Pick Stopped Being Yours

There's a second default stacked under the first.

Cursor sets the model to Auto by default, meaning it picks the model it considers best suited for each task rather than asking you, as that same dev.to tutorial notes. EastonDev's guide calls this the "daily default: Auto." So the defaults now cover three decisions at once: who writes the code (the agent), how it writes (autonomous multi-file). And with which brain (whatever Auto selects).

The models underneath are not exotic. Laxima's February 2026 guide says the pragmatic default for most engineering workloads is still Claude Sonnet 4.6, priced at $3/$15 per million tokens, with Claude Opus 4.8 and Cursor's Composer 2.5 as alternatives.

Composer is reportedly 4x faster than comparable models and completes most agent turns in under 30 seconds, per TechSifted.

The visible cost is the subscription: $20 Pro, $60 Pro+, $200 Ultra per that dev.to pricing table. The invisible cost is the token meter running behind Auto. This isn't a conspiracy, it's an incentive: the party picking your model is also the party billing your usage. Sonnet 4.6 at $3/$15 is a sane default for most work, which is exactly why nobody questions it. Defaults you never examine are the ones that compound.

The trade-offs are real too. And the honest comparisons say so. daily.dev's comparison table flags Claude Code for high token use with no IDE layer. And Cursor for context drift and sessions that stop when the app closes. Their recommendation is the one I'd echo: go with the tool that fits how your team already works.

Cursor for fast iteration inside the editor, terminal-native tools when the work is refactor-heavy.

The Real Bill Lands in Code Review

When the default becomes "agent writes," your default becomes "you review." That's the shift nobody put on a landing page.

Throughput stops being capped by typing speed and starts being capped by how many diffs you can genuinely evaluate in a day. For a one-person shop, review attention is the scarcest resource you own. And every vendor just got permission to spend it by default.

So audit your defaults this week, since they are decisions someone else made about your workflow:

- Open your editor's chat and check which mode it boots in. If it's Agent, you're delegating multi-file edits on every prompt, intended or not. - Check what Auto selected for your last few tasks, and whether that model's pricing fits the job. - Check whether the GitHub cloud agent is enabled on your client repos, and write down an actual yes or no. - Treat every agent-opened PR like a junior developer's PR: read the whole diff before you merge, every time.

None of this means ripping the agents out. Cursor and Copilot earn their keep on real work, and the front-runner list exists for a reason. The point is narrower and more useful: adoption already happened, silently, through defaults. The only move left to you is deciding whether those defaults serve how you ship or just serve the vendor's meter. That audit is a ten-minute job, and it's the cheapest engineering decision you'll make this quarter.

If you want a second pair of eyes on it, that's the work I do: Mediascout builds and audits AI coding workflows for small teams, defaults included.

Bring your setup and we'll find where the automation is actually pointed.