GitHub in the age of agents
Copilot grew from autocomplete into a team of agents, and GitHub's prices, data rules and uptime changed with it. What 2026 brought, and what it means for you.
For most companies that build software, GitHub is where the code lives, where it is reviewed and from where it ships. Copilot began as a suggestion engine inside the editor. Today GitHub presents its platform as a home for agents, and 2026 has shown what that means in practice: new ways of working, new prices, new rules for data, and a platform under strain.
From autocomplete to agents
The turn came in May 2025 with the Copilot coding agent. You assign an issue to Copilot, and it starts a virtual machine in GitHub Actions, works on the code and opens a draft pull request for review. The guardrails were sensible from the start: the agent can only push to branches it created, the person who asked for the work cannot approve it, and workflows do not run until a human approves them.
In October 2025 GitHub opened the platform to other makers' agents. Under the name Agent HQ, agents from Anthropic, OpenAI, Google, Cognition and xAI were to arrive inside paid Copilot subscriptions, steered from one mission control. GitHub said it had 180 million developers, and that 80 percent of new developers use Copilot in their first week.
In June 2026 came the Copilot app, a desktop app built around agents. Each session runs in its own copy of the branch, so parallel agents do not collide, and Agent Merge watches the checks, tracks the required reviewers and, if you allow it, merges. In October local sandboxing became generally available: commands an agent starts on a developer's machine run with restricted access to files, network and credentials, under rules an organisation can enforce.
A platform under load
Agents create work at a pace people never did, and GitHub felt it. In April its chief technology officer, Vlad Fedorov, wrote that GitHub had begun planning for ten times its capacity in October 2025, and by February 2026 was designing for 30 times, with agentic workflows as the main driver since late December. The same post apologised for a merge queue fault that produced incorrect merge commits in 2'092 pull requests across 658 repositories. No code was lost, but some default branches were left in the wrong state.
On 17 August GitHub was disrupted for 7 hours and 47 minutes, across sign in, Actions, the API, pull requests, issues and Copilot. A traffic peak met a component in one data centre that failed to scale. Commits had grown from 1.4 billion to 2.9 billion a month since April, Fedorov wrote, and he did not soften it: “We failed to scale critical components before demand exceeded their capacity.”
Prices and data
The cost of agents reached the price list. In April GitHub paused new sign ups for its Pro, Pro+ and Student plans, tightened limits and removed Anthropic's Opus models from Pro. Its reason: “it's now common for a handful of requests to incur costs that exceed the plan price.” A week later it announced usage based billing from 1 June. Every plan now includes credits worth its price, 10 dollars for Pro, 39 for Pro+, 19 per user for Business and 39 for Enterprise, and usage draws on them at each model's published API rates. Code completions stay included, and Copilot code review also uses Actions minutes.
In October GitHub let organisations bill code reviews to the company and restrict them to licences the company provides. Under the default setting, a review requested by a member whose quota is used up simply fails.
The data rules changed as well. Since 24 April, GitHub uses interaction data from Copilot Free, Pro and Pro+ users, including inputs, outputs and code snippets, to train its models unless the user opts out. Business and Enterprise users are not affected. That matters for any company whose developers use personal Copilot accounts on company code: private repositories are not used at rest, but code processed during a session can be.
What it means for a company
Five things follow. Put company code on Business or Enterprise licences, and keep personal plans away from it. Set budgets at enterprise, cost centre and user level before agents run, because work is now billed by the token. Keep branch protection and required reviews strict, because the agents' guardrails lean on them.
Plan for GitHub being down: if your releases, your agents and your reviews all run there, a long outage stops all three, so know how you would ship without it. And keep your instructions, skills and tests in the repository rather than in one vendor's agent, because the list of agents will keep changing.
That is how we build with Werk. AI writes most of the code on a framework we built and trust, people review every release, and every release is tested. GitHub gives agents a place to work. It does not decide what is good enough to ship. A person still does.
