GitHub launched GitHub Copilot Fusion at GitHub Universe on 28 October 2026. On the same day, GitHub's Octoverse report said that by September more than 40% of pull requests on GitHub started with an agent's first draft, up from 17% in March. If your team ships software, the question is no longer whether agents open pull requests. It is who answers for the merge. Name that person on every repository before you turn the volume up.
What is Copilot Fusion?
Copilot Fusion is a routing layer inside Copilot that decides which model, or which mix of models, handles each step of a coding task. In its Universe 2026 post, GitHub says Fusion "evaluates what a task needs across reasoning, code generation, debugging, and tool use," and then runs the work across one or more models.
The idea is sensible. Not every step needs the most expensive model. Renaming a variable and redesigning a data model are different jobs, and paying frontier prices for both is waste.
GitHub's headline claim, in the same post, is that Fusion Max scored within two points of Opus 5.5 Max on the Artificial Analysis Coding Agent Index, at half the cost and 40% less time per task. That is GitHub's number about its own product. Alongside has not tested it.
What else changed at Universe?
Two things matter for teams that buy Copilot for a company.
- Billing. GitHub says Enterprise customers can now use all Copilot features without buying a separate subscription. "You pay for the AI credits you use." It is available through GitHub Sales today, with self-serve coming later.
- Review. GitHub introduced Copilot code review Max and says it "posted the highest F1 score on ReviewBench for high-severity issues." Lite and Balanced remain for everyday reviews. GitHub's post does not say which plans include Max.
Paying for credits instead of seats changes the conversation with finance. A seat is a fixed cost. Credits move with how hard the agents work. If agents open more pull requests next quarter, the bill follows. Someone has to watch it, and someone has to decide which repositories deserve the expensive review mode.
How many pull requests do agents start now?
More than 40% by September 2026, according to the Octoverse 2026 report. GitHub puts the September figure at 110 million pull requests, about six times March's volume. The same report counts 1.3 billion merged pull requests over the past year.
GitHub is careful about one point, and so should we be. Its glossary calls the agent signal "a floor, since not all agent use may be detected." The real share is probably higher.
The figure measures who wrote the first draft. It says nothing about how many of those drafts were good, how many were merged, or how many were reverted a week later. That is the part your team still has to measure for itself.
It also tells you where the pressure goes. If agents write the first draft of almost half the work, the slow step in delivery is no longer typing. It is reading. Teams that do not plan for more review time will end up with a queue of tidy pull requests nobody has really looked at.
Who is accountable when an agent opens the pull request?
The person who merges it. Always.
This is my one strong opinion on all of this. The tool that opened the pull request is not accountable for the merge. Copilot cannot be paged at night. It cannot explain a decision to a client. It cannot be asked, six months later, why that migration dropped a column. A person can, and a person has to.
That sounds obvious. In practice it erodes. When agents open half the pull requests, reviewers start to skim. The diff looks tidy, the tests pass, the bot left a friendly summary. Approve. Next. The review becomes a ritual, and the merge button stops meaning "I stand behind this."
What should a team change?
- Name an owner per pull request. Not the agent, and not "the team." A person whose name is on the merge and who will answer for it.
- Label agent drafts. Make it obvious which pull requests started with an agent, so reviewers know to read the reasoning and not only the diff.
- Keep humans on the risky paths. Migrations, auth, billing and anything that touches personal data get a human author or a second human reviewer. Agents can still help. They do not get the last word.
- Use AI review as a filter, not a verdict. A strong automated review catches issues earlier. It does not replace the person who understands what the change is for.
- Track reverts. If agent-started pull requests get reverted more often, you will see it in a month. If they do not, you have earned the right to trust them a bit more.
Does model routing change how we pick tools?
A little. Routing inside Copilot means you choose less about which model runs, and more about what the agent may do: which repositories, which commands, which secrets, which branches. The model is becoming a setting. Permissions are becoming the product.
That is the same lesson we drew when we compared agent tools in our guide to the best AI coding agent harness in 2026. The software around the model decides most of what happens in your codebase. Fusion moves model choice into that software, which makes the permission rules matter even more.
What would I do this week?
Look at your last month of merged pull requests. Count how many started with an agent. Then look at who approved them, and how long each review took. If the approvals got faster while the volume went up, your review has become a ritual. Fix that before you buy more credits.
Where Alongside fits
Choosing how agents open pull requests, and who is allowed to merge them, is the kind of work covered by our technical advisory. If the concern is code already written by agents, a code audit provides a focused review of that code and its risks.