Contents

GitHub has made Copilot Fusion generally available and released Copilot Projects in public preview, adding automatic model selection and a shared workspace for organizing agent tasks. The practical decisions are where to delegate work, how much reasoning effort to use, and which changes should still receive human review.

Key Takeaways#

  • Copilot Fusion selects models automatically. Its Standard, High, and Max effort levels enable deeper analysis at the cost of higher spending and longer waits.
  • Copilot Projects brings reference materials, repositories, agent sessions, and artifacts together in a single workspace in the desktop Copilot app.
  • Shared Automations and the now generally available Agentic Workflows offer different ways to manage recurring repository tasks.
  • New review and security features expand automated checks, but some remain in preview and involve separate permissions or AI credit costs.

Fusion: Choose an Effort Level First#

Copilot Fusion is generally available in VS Code, the GitHub Copilot app, the CLI, and JetBrains IDEs. It combines automatic model selection with coordination across multiple models.

Standard uses one model for everyday tasks. High is intended for complex changes with a clear scope, while Max handles tasks requiring deep reasoning or broad analysis. High and Max use multiple models as needed.

Fusion decides whether to use a single model, pass an initial answer to a more capable model, or have other models critique an answer before revising it. These execution choices are automatic; users select the effort level rather than the coordination method.

| Effort Level | Recommended Use | Eligible Plans | |---|---|---| | Standard | Everyday changes | All Copilot plans | | High | Complex changes with a clear scope | Pro+, Max, Business, Enterprise | | Max | In-depth investigation | Pro+, Max, Business, Enterprise |

In the Copilot app, click the model name below the input field, enable Copilot Fusion, and select an effort level. In VS Code, select Fusion from the chat model menu.

Projects: Organize Work Around a Goal#

Now in public preview in the desktop app, Copilot Projects brings together related repositories, reference materials, sessions, and artifacts. It is useful when reaching a goal requires several related tasks, such as reviewing customer feedback and implementing improvements.

Create a project using the + next to Projects in the sidebar, then enter your goal. Add a local folder from the input field, or search for a GitHub repository by owner or repository name. A repository is optional.

Use Library to view attachments and generated artifacts, and Overview to track ongoing work, completed tasks, open pull requests, and decisions awaiting review. For a single, clearly defined bug fix, starting a session directly from a repository or folder is simpler.

Automate with Appropriate Controls#

Shared Copilot Automations belong to a repository, so recurring work can continue even after their creator leaves the team. People with read access can view the configuration, while those with write access can edit it. AI credits and Actions runtime are billed to the repository owner.

Automations can respond to events, external REST API calls, or the results of another Automation. Set a per-run credit limit before enabling them. Daily, weekly, and monthly limits can also stop new runs once the budget is reached.

Choose the now generally available Agentic Workflows if your automation definitions need Git history and pull request reviews. Commit both the Markdown definition and the generated .lock.yml file. Agents have read-only GitHub access by default. Validated writes run in a separate job through the safe outputs mechanism.

Check Before Enabling Approvals#

Max is coming to Copilot code review in public preview. Multiple agents independently review changes using models from different providers, and another agent consolidates their findings. GitHub positions it for changes to authentication code, shared libraries, and similar areas. It takes longer and consumes more credits than lighter review settings.

Copilot approvals are generally available but disabled by default. Administrators can allow Copilot to approve pull requests and configure those approvals to count toward merge requirements. New commits invalidate the approval and require another review.

Agent Merge can monitor pull requests for Actions failures, merge conflicts, or both. If GitHub auto-merge is enabled, repaired pull requests can merge once they pass the required checks and receive the necessary approvals. Review those requirements before combining automatic repairs with automatic merging.

Check Costs and Preview Limitations#

Fusion itself carries no extra charge, but usage is billed based on the models invoked. Tasks using multiple models may consume more AI credits. Paid plans receive a 30% discount on model costs incurred through Fusion until November 30, 2026.

According to GitHub, Fusion Max scored 60 on the Artificial Analysis Coding Agent Index, compared with 62 for Claude Code using Opus 5.5 Max, at roughly half the average API cost per task. This comparison evaluates specific model and agent configurations; it does not establish which will deliver the best results in your repository.

There are other points to keep in mind when adopting these features.

  • Cloud sandbox compute moves to usage-based billing in November 2026, with eligible free Actions allowances available to offset usage.
  • The preview Security extension in the Copilot app can review local changes on all plans. Scan alerts for private repositories have additional eligibility requirements.
  • Dependabot reachability analysis is in preview for JavaScript and TypeScript on Enterprise Cloud. Its findings do not automatically dismiss alerts.
  • Fusion’s experimental routing to local models also requires an internet connection because routing decisions happen in the cloud.

Next Steps#

Run one representative repository task with Fusion Standard, then review the diff, test results, and AI credit usage before choosing a higher effort level.

Source: https://gihyo.jp/article/2026/10/github-universe-2026

Contents