GitHub Copilot's Coding Agent: Three Surfaces, Three Defaults

GitHub Copilot’s coding agent is the cloud one. You assign it a GitHub issue, it works on its own inside a firewalled, ephemeral sandbox, and it opens a pull request for you to review. That is a different product from Copilot’s agent mode in VS Code, and different again from Copilot CLI. All three are Copilot, all three are agents, and searching for “GitHub Copilot coding agents” will return documentation for whichever one the page you land on happens to mean.

The distinction is not pedantic. The three surfaces have three different answers to the question that actually determines what you can safely point them at: does anything stop this before it acts? And they draw on the same pool of paid credits, so the one you use changes what the bill looks like.

The three surfaces, and what each does before it acts

We read GitHub’s own permissions and approvals documentation for all three on 11 August 2026, as part of our census of autonomy defaults across 21 coding agents. This is what we found.

Surface	What it is	Default before it edits a file or runs a command	How you turn the gate off
Agent mode (VS Code)	Copilot working inside your editor, on your machine	“Default Approvals”: a confirmation dialog before any tool that isn’t pre-classified as safe	Bypass Approvals
Copilot CLI	Copilot in your terminal	Auto-approves read-only actions only; anything that can modify the system prompts every time	--allow-all / --yolo
Copilot coding agent (cloud)	Assigned to a GitHub issue, runs on GitHub’s infrastructure	No per-action human gate at all. It runs unsupervised inside a firewalled, ephemeral sandbox	Not applicable; see below

The third row is the one people mean when they search for this, and it is the one that behaves least like the other two. The approval checkpoint does not disappear; it moves downstream. Instead of clicking yes before each command, you are deciding whether the resulting pull request’s CI workflow is allowed to run at all. If you are used to agent mode’s dialog box, the cloud agent will feel like a different tool, because on this axis it is.

That matters for a second reason: averaging these into one row would misrepresent all three. Anywhere you see “GitHub Copilot: asks first” or “GitHub Copilot: fully autonomous” stated flatly, one of the three surfaces is being reported as if it were the product.

What it costs, and why the answer is “credits”

Copilot’s published plans, recorded from GitHub’s own pricing page and last verified on 9 August 2026:

Plan	Price	What the published feature list says
Free	Free	2,000 completions/month + 50 chat requests (incl. Copilot Edits)
Pro	$10/month	Unlimited completions; includes $15/mo in GitHub AI Credits for chat, agent mode, code review and the cloud agent
Pro+	$39/month	Unlimited completions; includes $70/mo in AI Credits; premium models incl. Opus
Max	$100/month	Unlimited completions; includes $100/mo in AI Credits; priority model access
Business	$19/user/mo	Org policy and controls, pooled usage
Enterprise	$39/user/mo	2× Business’s usage; priority support add-on

The important line is the Pro row. Agent work is not billed as a seat, it is billed out of a credit allowance, and the same allowance covers chat, agent mode, code review and the cloud agent together. So the honest answer to “what does the Copilot coding agent cost” is $10/month to get access to it, and then it depends entirely on how much of the $15 credit pool your agent runs consume against everything else drawing on that pool. The published plan price is a floor, not a forecast.

One absence worth stating as an absence: the Free tier’s published allowance lists completions and chat requests, and does not name AI Credits or the cloud agent. We have not verified whether the cloud agent is reachable on Free, only that GitHub’s own feature list doesn’t say it is.

For context, our entry-price census of all 81 tools in the index counted 14 coding agents publishing an exact flat monthly price, ranging from $8 to $25 with a median of $18.50. Copilot Pro at $10 sits near the bottom of that band. The nearest published prices in our own data:

Tool	Cheapest paid plan
OpenAI Codex CLI (Go)	$8/month
GitHub Copilot (Pro)	$10/month
Junie (AI Pro, individual)	$10/month
Claude Code (Pro)	$17/month billed annually, $20 monthly
Cursor · Devin · Windsurf · Factory · Replit Agent (Pro/Core)	$20/month

Where Copilot sits among the 21 coding agents we index

These are the five attributes we record identically for every tool in the coding agents category, so the comparison is exact rather than impressionistic.

What we record	GitHub Copilot	How common that is among the 21
Interface	IDE	9 are CLI-first, 6 IDE, 5 Web, 1 visual
Deployment	Cloud	8 Cloud-only, 8 both, 5 self-host-only
Open source	No	12 closed, 9 open source
Free tier	Yes	18 of 21 have one
Models	GPT, Claude and Gemini, selectable	not counted

Two things follow that are worth knowing before you commit.

Copilot is cloud-only in our record. Five of the 21 coding agents we index run entirely on your own machine and eight offer both. If the constraint you are working under is that source code cannot leave your infrastructure, that is the column to filter on first, and it rules Copilot out before any of the autonomy or pricing questions matter.

On training defaults, Copilot changed sides this year. Per our census of what coding agents do with your code, GitHub flipped its default on 24 April 2026: interaction data from Free, Pro and Pro+ accounts now trains Copilot’s models unless you opt out in account settings, a reversal of the prior opt-in model. Business and Enterprise seats never train. Copilot is not alone here: 8 of the 21 tools we checked train by default on their individual tier, and Claude Code went through the same reversal eight months earlier. But if you evaluated Copilot before April and are relying on what you concluded then, that specific conclusion has expired.

What we have not checked, and are not going to imply

The question behind this search is often really “which other coding agents will take an issue and hand me back a pull request without supervision?” We have not measured that dimension across the index, and an absence in our data is not a zero. Our records cover interface, deployment, licence, pricing and models for all 21 tools, and we have separately checked autonomy defaults, training defaults and Terminal-Bench scores. But “offers an asynchronous cloud agent you assign work to” is not a field we have researched tool by tool, and we are not going to infer it from the deployment column, because “Cloud” and “runs unattended on a ticket” are not the same claim.

Similarly: Copilot does not appear on Terminal-Bench, which we scored across the category on 2 August 2026. That is a statement about the benchmark’s coverage, not about the tool. Terminal-Bench measures autonomous terminal task completion, which is not the shape of an IDE-embedded assistant, and reading a missing row as a low score would be wrong.

If you’re deciding today

Work out which of the three surfaces you actually want first, because the rest follows from it. If you want the editor experience with a confirmation step, that is agent mode, and its default already does what most people assume the whole product does. If you want to hand off a ticket and come back to a pull request, that is the cloud coding agent, and you should treat its CI permissions as the real security boundary, since there is no per-action prompt standing in front of it. Either way the entry price is $10/month, the meter that matters is the credit allowance rather than the seat, and if your code cannot leave your own machines, five of the tools in this category can run there and Copilot is not one of them.

Methodology: Copilot’s plan names, prices and published feature lists are recorded in our GitHub Copilot listing from GitHub’s own pricing page, last verified 9 August 2026. The autonomy defaults for all three surfaces come from GitHub’s own approvals and responsible-use documentation, read directly on 11 August 2026 and cited in full in our autonomy defaults census: principally GitHub’s responsible-use page for the coding agent and the VS Code approvals documentation. The training-default finding is sourced in our data-training census. Category-wide counts are computed from the 21 published listings in coding agents as of 13 August 2026. Vendor defaults and prices change, sometimes on a published date, so check the primary source before relying on any of this for a purchasing or security decision.