[3/3] Octoverse 2026: Agents, Models, and Development Workflows

To wrap up this series on Octoverse 2026 (parts 1 and 2), let's look at what happens between defining a task and merging a PR. This is where agents, model selection, and code review come together.

🤖 Agents: from asking a question to delegating a task

According to GitHub, more than 40% of PRs in September 2026 started with a detected agent-generated draft, up from 17% in March. In September that represented 110 million PRs on the platform. Among Copilot's monthly active users, 55% used agent features, up from 35% a year earlier. An agent draft doesn't mean an agent worked autonomously on the whole PR. But the beginning of the task already looks different: a developer describes what needs to be done and gets a change to discuss, check, and refine.

Here's a striking figure: new repositories with an agent configuration had eight times as many commits on average as new repositories without one. Those groups may already have differed in their teams and tasks, and the absence of a configuration file doesn't rule out AI use. The comparison doesn't establish an eightfold gain in productivity.

🧠 Models: several tools in use

In September, 61% of Copilot users who chose a model manually used more than one across different sessions. Paying Copilot users used 47 different models that month, and no model accounted for more than 25% of requests. The request sample excludes automatic selection, requests using a user's own API key, and service traffic. This describes model usage within Copilot. It tells us neither overall market share nor which models are better. GitHub separately reports that 63% of paid Copilot users used automatic model selection at least once in September. The same person can use both manual selection and automatic routing.

I'm interested in matching the model to the stage of work. Planning, implementation, and verification can have different requirements for quality, speed, and cost.

🔎 Review and security: what happens after generation

GitHub says 90% of agent-assisted PRs in September merged within two weeks, versus about 86% of PRs without detected agent involvement. In February the figures were 83% and 86%, respectively. These data don't tell us about the quality of changes after release. The tasks may have differed and the group without detected agents may still use AI.

Use of AI-assisted code review and GitHub Advanced Security grew more than 90% year over year for each tool. Verification tools are getting more attention too, although greater use doesn't establish that there are fewer defects.

A team can measure a change from task definition through release: how long the agent, review, and rework took, what it cost, and whether the result had to be rolled back. That gives you a basis for deciding where delegation helps and which parts of the process are worth changing.

#Engineering #AI4SDLC #Agents #DevTools #Metrics #Security