--- title: "GitHub Universe 2026: 3 features to try now, terms and costs" description: "Agentic Workflows, Copilot app Projects, and Copilot Dreaming, checked against GitHub docs: release stage, plans, how AI credits are charged, and what changes without approval." url: https://udon-dev.com/en/notes/github-universe-2026-try-today/ lang: en published: 2026-11-03 topic: [ai] tags: ["github", "github-copilot", "ai-agent", "github-actions"] --- # GitHub Universe 2026: 3 features to try now, terms and costs - Agentic Workflows became generally available on October 28. It compiles instructions written in Markdown into an Actions workflow and runs it, and writes are limited to the operations declared in the frontmatter - Inference per run is capped by default at 1,000 AI credits (1 credit = $0.01), and the cap can be changed with `max-ai-credits` - Projects in the Copilot app is a public preview available on all Copilot plans; you get it by updating the app. It lists work across multiple repositories by status, such as "Needs input" and "Running" - Copilot Dreaming is an opt-in public preview, not available on Copilot Pro. Memories are added and deleted without approval, and changes to files like CLAUDE.md and AGENTS.md are proposed as PRs. Runs use AI credits even when nothing changes GitHub Universe 2026 took place in San Francisco on October 28 and 29. The keynote announcements ranged widely, from Copilot Fusion (automatic model selection) to the Copilot app, Projects, Automations, and Max for code review, more than anyone can follow all at once. An engineer at one of GitHub's partner companies attended in person and wrote a report on [the company's tech blog, "Alternative Architecture DOJO"](https://aadojo.alterbooth.com/entry/2026/11/01/230000) (in Japanese). It names three things as "usable starting today": GitHub Agentic Workflows, Projects in the Copilot app, and Copilot Memory with Copilot Dreaming. The author also wrote of wanting to replace the work notes they had been keeping in a local `OREORE-FEATURES.md` ("my own features") with Projects, Memory, and Dreaming. But what "starting today" means differs a lot by feature. Is it generally available or in preview? Which plans include it? What costs AI credits, and what changes without a person's approval? I checked these against GitHub's official documentation and Changelog. ## Availability of the three features | Feature | Stage | Who can use it | How it is charged | |---|---|---|---| | Agentic Workflows | Generally available (October 28) | GitHub Free, Team, Enterprise. The default Copilot engine requires a Copilot license | Actions run time, plus AI credits for inference | | Projects in the Copilot app | Public preview | All Copilot plans | Not stated in the Projects documentation | | Copilot Dreaming | Public preview, opt-in | Eligible paid plans. Not Copilot Pro | AI credits per run | Copilot Memory itself became [generally available](https://github.blog/changelog/2026-10-28-github-intelligence-puts-your-code-context-and-history-to-work) on paid Copilot plans the same day. Dreaming is built on top of Memory. Also, from October 28, GitHub Enterprise customers can use all Copilot features without a separate subscription ([GitHub's announcement](https://github.blog/news-insights/company-news/github-universe-2026-evolving-every-layer-of-the-stack/)). For now it is offered through sales representatives, with self-service sign-up coming soon. Pricing is based on AI credit usage, and existing Copilot Business and Copilot Enterprise customers can choose to stay on their contract or move. ## Agentic Workflows: declare writable operations up front [GitHub Agentic Workflows](https://docs.github.com/ja/copilot/concepts/agent-capabilities/agent-powered-workflows-and-operations/about-github-agentic-workflows) lets you hand repetitive work to a coding agent running on GitHub Actions. Examples include issue triage, investigating CI failures, updating documentation, and daily reports. It started as open source from GitHub Next and became [generally available on October 28](https://github.blog/changelog/2026-10-28-github-agentic-workflows-generally-available). A workflow is a single file made of YAML frontmatter and a Markdown body. The frontmatter holds the trigger, repository permissions, and the write operations allowed (safe-outputs); the body holds natural-language instructions for the agent. You compile this file into an Actions workflow, `.lock.yml`, and commit both. Because it runs as ordinary Actions, your existing runner groups and policy constraints still apply. （図: Image from the GitHub Changelog. Under the heading "GitHub Agentic Workflows is now generally available," a diagram shows a run of daily-repo-status.lock.yml progressing through activation, agent, detection, and safe_outputs） Source: [GitHub Agentic Workflows generally available (GitHub Changelog)](https://github.blog/changelog/2026-10-28-github-agentic-workflows-generally-available) This is the part where the event report stressed guardrails. The mechanisms the documentation lists are: - Unless stated otherwise, repository permissions are read-only - Writes such as creating issues or comments and creating PRs go through the safe-outputs declared in the frontmatter - Secrets are not passed to the agent's execution environment; they sit in a separate job - Before anything is written, a dedicated threat detection job inspects the output - The agent runs in an isolated container behind a firewall The default engine is Copilot, and Claude, Codex, and Gemini can also be chosen. With engines other than Copilot, inference is billed by that provider. [Repo Assist](https://dsyme.net/2026/02/25/repo-assist-a-repository-assistant/), which the event report introduced, is an example of using these constraints. According to its author, Don Syme, it can only apply labels from a fixed set, and it can add comments but not delete them. The number of PRs it can open in one run is limited, and their titles get a fixed prefix. It can only add commits to PRs it created itself. He says he built it to clear out issues in neglected repositories. ### The cost cap Inference cost is estimated in AI credits (AIC). 1 AIC is $0.01. The default cap per run is 1,000 AIC, so with no settings a run stops at $10 worth. You can change the cap with `max-ai-credits` in the frontmatter. To check after a run, use the GitHub CLI extension. `gh aw logs` shows estimated time, tokens, and AIC for recent runs, and `gh aw audit ` shows details for one run. With general availability, `gh aw forecast`, which projects monthly spend, also left its experimental stage. However, the documentation calls AIC values best-effort estimates and says to confirm final amounts with the provider's billing. ## Projects in the Copilot app: group work across repositories under one goal [Projects](https://docs.github.com/ja/copilot/how-tos/github-copilot-app/working-with-projects) are goal-based bundles of work that you create from the Copilot app's sidebar. You can add several repositories, or start without any. When you write the goal and first task in the opening prompt, Copilot runs as many agents as needed and returns results such as PRs. What the event report said "got easier to use" is the screen for following progress. In the "Overview" on the right, a project's work is split into four groups: "Needs input," "Open," "Running," and "Done." Decisions, reviews, and permission requests gather under "Needs input," and open PRs are listed under "Open" with lines added and removed. It is built so you can see where things are waiting on your reply without opening each session. There are also tabs for a "Library," which holds materials such as specs and customer feedback, and "Automations," which manages scheduled work. If you ask in conversation, for example "summarize feedback every Monday at 9:00," a scheduled run is set up, and you can pause or delete it from the tab. According to the [Changelog](https://github.blog/changelog/2026-10-28-github-copilot-app-universe-releases), information from connected services such as Slack and Figma can also be used for research and drafting specs. The same day brought Computer use, which operates desktop apps (macOS and Windows, public preview). It asks for approval before operating an app, and organizations can disable it in admin settings. According to the event report, Projects is available once you update the Copilot app to the latest version. The Copilot app itself is available on all Copilot plans, so of the three, this looks the easiest to try. ## Copilot Dreaming: memories change without approval, instruction files arrive as PRs [Copilot Dreaming](https://docs.github.com/ja/copilot/concepts/prompting-and-context/copilot-dreaming) analyzes past sessions and reactions to PR reviews in the background, and keeps Copilot's memories and the repository's instruction files up to date. It updates the context given to Copilot, not the model itself. There are three kinds. | Kind | Information used | Result | |---|---|---| | Repository customization | Repository sessions, code, and memories | Updates to repository memories, and proposed changes to instruction files and skills | | Personal memory consolidation | Your own sessions and personal memories | Updates to personal memories | | Code review customization | Review reactions on merged PRs | Guidance for Copilot code review | What is worth checking is that changes land differently for each kind. Memories are added, updated, and deleted without per-change approval. Changes to files such as `.github/copilot-instructions.md`, `AGENTS.md`, `CLAUDE.md`, `GEMINI.md`, and `.github/skills/` are proposed as PRs, and the files do not change until you merge. Code review guidance is updated internally without going through a PR. Since `CLAUDE.md` and `AGENTS.md` are included, instruction files you feed to Claude Code or Codex will also get change PRs from Copilot. The documentation warns that AI-generated guidance may be wrong and says to check it before merging. ### Prerequisites and cost In the early preview, GPT-6.1 Sol must be allowed by policy, and it is not available on Copilot Pro. It does not support GitHub Enterprise Cloud with data residency either. In an organization, an administrator enables it by policy, and for personal Dreaming the user also enables it in their own settings. Cost is billed as part of existing AI credit usage, with no separate line item for Dreaming. According to the documentation, runs that make no changes, or that process inactive repositories, still use credits. It says to set a budget before rolling it out across an organization, and states plainly that alerts alone do not stop usage. Disabling Dreaming does not revert changes, close the PRs it opened, or restore deleted memories. You can review and delete memories in personal or repository settings. ## What the event report ended with After covering the features, the event report closes on team culture. Designate someone (a champion) you can ask when you don't understand something, build an environment where people can ask, and aim for everyone to move up to the same level. The author says this point, long made in common across sessions about rolling out Copilot, was repeated again this time. If you try these as a team, it seems better to start with something like Projects that you get by updating the app, and to enable Agentic Workflows and Dreaming only after setting an AI credit budget.