The year AI agents became coding partners

If 2024 was the year of AI models, 2025 was the year they started doing real work. Across the most-read posts on the GitHub Blog this year, one theme dominated: AI agents that don't just suggest code, but plan, execute, and fix their own mistakes. Here's a look at what developers engaged with most over the past twelve months.

From agent mode to the coding agent

GitHub's agent mode arrived in February, bringing a new level of autonomy to Copilot inside the IDE. Rather than waiting for prompts, the agent can iterate on its own code, recognise errors, and correct them in real time. It functions as a problem solver: it takes your intent, builds a solution, and keeps refining until the result is right.

Later, in May, the coding agent moved that capability directly into GitHub. Assign an issue to Copilot or prompt it from VS Code, and the agent will explore the repository, write code, run tests, and open a pull request for your review. The workflow shifts from "write code" to "review work."

By October's GitHub Universe, the picture had widened further. Agent HQ was announced as a platform for unifying agents from multiple vendors—Anthropic, OpenAI, Google, Cognition, xAI, and others—so they can work together inside GitHub as part of a paid Copilot subscription.

MCP becomes the glue

Model Context Protocol (MCP) emerged as the standard way for agents and tools to communicate. Each MCP server functions like an ingredient in your AI stack, whether you're using Playwright for browser automation, Notion for knowledge retrieval, or GitHub's own server with more than a hundred tools.

The rollout was incremental but steady. In April, agent mode with MCP support became available to all VS Code users. By September, the GitHub MCP Registry launched as a central place to discover, build, and evaluate MCP servers—removing much of the friction in finding the right one for a given task.

Spec-driven development

One notable shift in methodology was the rise of spec-driven development. The idea is simple: instead of coding first and documenting later, you write specifications that serve as the shared source of truth throughout the project. To support this workflow, Spec Kit—an open source toolkit from GitHub—provides a structured process for integrating specifications into coding agent workflows, putting the spec at the centre of engineering.

Conversations that shaped the year

Beyond the product announcements, several long-form pieces stood out. Git turned 20 in 2025, and the blog marked the occasion with a Q&A with Linus Torvalds, its creator and the mind behind Linux. There was also a deep dive into Log4Shell, the 2021 vulnerability that became a defining moment for open source security. The interview with maintainer Christian Grobmeier offered a rarely heard perspective on the breach that broke the internet. Closer to home, a conversation with Home Assistant maintainer Franck Nijhof explored how the local-first project became a fixture in millions of households—and the paradox that makes it so appealing to developers.