What did GitHub announce on October 28, 2026?

GitHub published five changelog entries: a connected Copilot cloud experience with cloud sandboxes, new triggers, tools and credit limits for Copilot Automations, Copilot Fusion (a model router), Agent Merge for pull requests, and the general availability of GitHub Agentic Workflows.

What are Copilot cloud sessions and sandboxes?

Per GitHub, one session can be started and continued across github.com, GitHub Mobile, the Copilot CLI, the Copilot app, VS Code, Slack and Teams. Each session runs in a cloud sandbox that saves a snapshot on shutdown and has no one-hour limit on total running time. It is generally available on paid Copilot plans.

What is Copilot Fusion, and is it related to Oracle Fusion?

Copilot Fusion is GitHub's router that picks and combines models for each task at Standard, High or Max effort. It has no connection to Oracle Fusion Claw, which we cover separately. It is generally available in VS Code, the Copilot app, the CLI and JetBrains IDEs.

What does Agent Merge do?

According to GitHub, Agent Merge watches a pull request and works in a cloud environment to fix failing Actions workflows and merge conflicts, and it can be combined with auto-merge. It is available on paid Copilot plans, subject to organization and enterprise policies.

Is GitHub Agentic Workflows generally available, and for whom?

GitHub says it is generally available on github.com for Free, Team and Enterprise. Workflows are written in Markdown and compiled to Actions YAML. Using the default Copilot CLI engine requires a Copilot license.

How is cloud sandbox usage billed?

GitHub says billing for cloud sandbox usage begins in November 2026 and draws from the included Actions entitlement, now one shared dollar-based pool. The changelog gives no sandbox price, and GitHub's documentation may publish rates.

How much does Copilot Fusion cost?

GitHub's documentation says Fusion has no separate charge: you pay each model it uses at that model's standard rate. Paid plans get a 30% discount on model costs through November 30, 2026. One AI credit equals $0.01, and Fusion High and Max may use more credits than Standard.

What should a team review first?

Our recommendation: credit limits on every automation, which events may start an agent, who owns the merge, the data and MCP policies, and a test of Fusion's cost per completed task on your own work. We have not tested these products.

GitHub Copilot Cloud Agents: Fusion, Merge & Costs

A row of cloud sandbox blocks joined by thin lines to a pull request node, with one red line marking the merge step

The short version

On October 28, 2026, GitHub published five changelog entries about running Copilot agents in the cloud: a connected cloud experience with cloud sandboxes (generally available on paid plans), Copilot Automations with 24 new triggers, 20 new tools and AI credit limits, Copilot Fusion (a model router, generally available), Agent Merge for pull requests (paid plans) and GitHub Agentic Workflows (generally available on Free, Team and Enterprise). GitHub says billing for cloud sandbox usage begins in November 2026 from a shared dollar-based Actions entitlement, and the changelog gives no sandbox prices. GitHub's documentation says Fusion has no separate charge and that paid plans get a 30% discount on model costs through November 30, 2026. GitHub also says Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task. All figures are GitHub-reported, and we have not tested any of these products.

On October 28, 2026, GitHub published five changelog entries about running Copilot agents in its cloud, tagged Universe 2026. They cover where a coding agent can run, what can start it, how it picks a model, how it gets a pull request merged and how compute is billed. They are related announcements, not one product with one execution environment. This teardown goes through those five documents, plus two pages from GitHub's own documentation for billing details. It is not a hands-on test: every claim below is GitHub's, the arithmetic and recommendations are ours, and we have not tested any of these products. For how Copilot compares with Claude Code, Cursor and Codex as tools, see our coding agents comparison; this page covers only what GitHub published on October 28.

What did GitHub announce on October 28?

Piece	What GitHub says it is	Availability stated	Not stated in the entry
Connected Copilot cloud experience and cloud sandboxes	One session you can start and continue across github.com, GitHub Mobile, the Copilot CLI, the Copilot app, VS Code, Slack and Microsoft Teams. Every session runs in a cloud sandbox that saves a snapshot on shutdown and has "no one-hour limit on total running session time".	Generally available on paid Copilot plans, rolling out from October 28. Copilot Free users keep Chat on github.com and GitHub Mobile.	Prices for sandbox usage.
Copilot Automations	Prompts, triggers and tools combined into workflows. New: a REST API trigger, 24 more GitHub event triggers, chaining of automations, 20 new GitHub tools, custom MCP server tools, and AI credit limits.	Copilot Pro, Pro+, Max, Business and Enterprise, in any repository where you have write access.	What a credit costs, and whether the new features are preview or general availability.
Copilot Fusion	A router that picks and combines models for each task, at an effort level you choose: Standard, High or Max. It works in three ways: Single, Cascade (a cheaper model drafts, a quality check decides whether to escalate) and Critique (one model drafts, another reviews, the first revises).	Generally available on all Copilot plans in VS Code, the Copilot app, the Copilot CLI and JetBrains IDEs. GitHub's Fusion documentation adds that its plan table does not list High and Max for the Free, Student or Pro plans, so check your plan.	The list of models it can choose from.
Agent Merge	"Watch and fix" in the pull request merge box: Copilot works in a cloud environment to fix failing Actions workflows and merge conflicts, and can be combined with auto-merge.	Paid Copilot plans, subject to organization and enterprise policies.	Quotas or costs per fix.
GitHub Agentic Workflows	An open source project from GitHub Next. You write a workflow in Markdown, and it compiles to standard Actions YAML so coding agents can do tasks such as issue triage or CI failure analysis inside GitHub Actions.	Generally available on github.com for Free, Team and Enterprise. The default Copilot CLI engine needs a Copilot license.	Prices.

How is the compute billed, and what does it cost?

GitHub says the included Actions compute is now one shared, dollar-based entitlement instead of a fixed number of standard runner minutes. It covers standard GitHub-hosted runners, cloud sandboxes and, on Team and Enterprise Cloud, larger hosted runners. Each product keeps its own published rate, so higher-rate compute uses the entitlement faster, the entitlement resets each billing cycle, and usage beyond it follows your existing rates, budgets and payment settings. GitHub says billing for cloud sandbox usage begins in November 2026. The changelog gives no sandbox price, so we cannot estimate a cost per session; GitHub's documentation may publish rates, so check it before you budget.

The model work is a separate meter. GitHub's usage-based billing documentation says 1 AI credit equals $0.01 and lists the Copilot cloud agent among the features that consume credits. Copilot Fusion has no separate charge: you pay each model it uses at that model's standard rate, and its documentation says paid plans get a 30% discount on model costs through November 30, 2026. Fusion High and Max may use more credits than Standard because they can call several models. The credit limits on Automations are set in AI credits, so they limit model spend, not sandbox compute.

What controls does GitHub describe?

Control	What GitHub says
Credit limits on Automations	Optional limits per run, day, week or month. A run stops when it reaches the per-run limit. Once a daily, weekly or monthly limit is reached, new runs do not start until the period resets. Automations without limits keep running as before.
Who can trigger an automation	A trigger's "Can be triggered by" setting defaults to "Only users". Setting it to "Users and Copilot" lets events from other automations, such as a pull request Copilot opened, start the trigger. Two new triggers run one automation after another succeeds or fails.
Agent permissions	In cloud sessions Copilot asks for approval before any tool call that requires it. Enterprise-managed settings for agent permissions and MCP server allowlists apply to the new cloud experience.
Data scope	Chat sessions are scoped to a single organization or your personal repositories by default. Organization members use single sign-on to reach session data. Without write access to a repository, Copilot use there is read-only.
Agentic Workflows safeguards	Read-only permissions by default, a sandboxed container behind the Agent Workflow Firewall, validated outputs ("safe outputs"), and a threat-detection job that scans proposed changes before they are applied.
Fusion administration	Fusion only uses models enabled in your enterprise and organization policies, and enterprise administrators can disable models for Fusion without disabling them for everyone, or set Fusion as the default.

What does GitHub claim about Fusion, and who measured it?

GitHub says that on the Artificial Analysis Coding Agent Index, Copilot Fusion Max scored 60, within two points of Claude Code with Opus 5.5 Max, at about half the average API cost per task, using about half as many tokens and taking about 40% less time per task. We read Artificial Analysis's public leaderboard on October 30. It lists Claude Code with Opus 5.5 (Max) at 62 and about $11.42 average API cost per task, the comparison point GitHub cites, but we did not find a Copilot Fusion entry on the pages we read, so the 60 is GitHub's own statement for now. The changelog does not describe the test setup, and the index averages three benchmarks across entries that run in different agent tools. The entry also carries a quote from an AT&T engineer saying Fusion Standard improved token efficiency, and the Agentic Workflows entry carries a quote from a senior manager at Accenture, who says one lead saved an estimated 29 hours a week, that AI authors 91% of the team's commits, and that production errors fell from about 11,000 a day to under 1,000. Both quotes are testimonials hosted on GitHub's own pages, not independent measurements.

What should engineering and operations teams decide?

These are our recommendations, not GitHub's.

Set budgets before you switch things on. Put a per-run credit limit on every automation, and daily or monthly caps where you can. Remember that a per-run limit stops work midway, while the other limits only block new runs.
Decide which events may start an agent. Leave "Can be triggered by" on "Only users" unless you want agent-opened pull requests to start more automations, and review any chain of automations as one system.
Decide who reviews and who authorizes the merge. Agent Merge plus auto-merge means an agent can fix a blocker and GitHub can merge once the requirements are met. Decide who reviews agent-generated fixes, who authorizes merging and whether auto-merge is allowed, keep your required reviews and branch protections, and note that automations can be given a tool that merges pull requests.
Test the router's savings on your own work. Run a fixed set of your tasks through Fusion and through the model you use now, and compare cost per completed task, not price per token. The 30% discount runs through November 30, 2026, so test at the price you will pay afterwards.
Check the data boundary. Review the "Copilot in the cloud" policy, the MCP allowlists and which repositories are opted out before agents get access.
Plan for November. Sandbox billing starts then, so look at how much of your included Actions entitlement agents will draw, and keep a separate budget for AI credits.

What does this mean if you run a business, not a codebase?

AI agents that work on code in the background are becoming a standard way to build software. For an owner, that means two questions to ask whoever builds or maintains your software: who reviews what the agent changed, and what it costs. We have not tested these products, and this section rests on GitHub's announcement.

What is still unknown?

The changelog entries give no price for sandboxes and no preview or general-availability label for the Automations changes. Fusion's documentation does not list the cloud models it routes between; it says the pool changes over time and names only the two local models of a separate public-preview feature. The benchmark figure and the testimonials are GitHub's own. Separately, researchers reported on October 29 flaws in the embedded browsers of several AI coding assistants, including a Copilot address-spoofing issue that they say GitHub fixed on October 13; see our EMBRACE decode. For governance questions across agent platforms, see our enterprise agent platform comparison, and for the day's other items see our October 30 digest.

GitHub Copilot Cloud Agents: What GitHub Changed on October 28, 2026