May 2025, Revisited: GitHub Copilot's Agent Moment

This post is part of our Three-Year Retrospective series: thirty-six posts, one per month, looking back at what actually mattered in software engineering. This one covers May 2025 — and closes out the second year of the series.

At Build 2025, on May 19, Microsoft made its answer to the agentic coding wave official: the GitHub Copilot coding agent. The pitch fit in one sentence and still landed strangely on first hearing — you assign a GitHub issue to Copilot, the way you'd assign it to a teammate, and the agent goes off and does the work. It spins up an ephemeral development environment in GitHub Actions, reads the codebase, makes its changes, runs the tests, and comes back with a draft pull request, pinging you for review like any other contributor. Alongside it, agent mode — the autonomous multi-step editing experience in VS Code — completed its rollout to general availability, and Microsoft announced it was open-sourcing the Copilot Chat experience in VS Code itself.

From autocomplete to teammate in four years

The distance traveled deserved a moment's appreciation. Copilot launched in 2021 as gray-text autocomplete — brilliant, narrow, and entirely synchronous: it worked exactly when your cursor did. Chat arrived, then multi-file edits, then agent mode through early 2025. The coding agent was the endpoint of that arc: asynchronous software work. You delegate at 4 p.m., review a PR the next morning. The unit of interaction stopped being the keystroke or even the conversation, and became the task.

Just as telling was where the agent lived. Not in the editor — in the repository. Its workspace was CI infrastructure, its identity was a collaborator on the repo, its output was a pull request, and its guardrails were the ones the platform already had: branch protections, required reviews, the firm rule that the agent's PRs needed human approval and couldn't approve themselves. The whole design said: agents are contributors, so put them where contributors are governed.

Microsoft's structural bet

That placement was the strategic content of Build 2025. Cursor's bet was the editor; Anthropic's, three months earlier, was the terminal. Microsoft, owning GitHub, bet on the repo — reasoning that whoever owns the place where code is reviewed, merged, and governed owns the natural habitat for software agents. Issues become task specs. PRs become the agent's deliverable format. Actions becomes the agent's compute. Notably, the announcements leaned into openness at the edges — MCP support threaded through the tooling, the VS Code AI experience going open source — a tacit acknowledgment that the protocols, at least, had already been standardized elsewhere.

For engineering teams, the questions this raised in May 2025 were brand new and instantly practical. Which issues are well-specified enough to hand to an agent? (The honest answer that spring: well-scoped, well-tested chores — and writing issues that clearly turned out to be a skill most backlogs failed.) Who reviews agent PRs, and how carefully? Does an agent's contribution graph count toward anything? Half of our teaching that summer was less about the tool and more about the management discipline around it.

What it was honestly like at launch

Early reports matched our own experience: the coding agent was genuinely useful on the tasks it was designed for — dependency bumps, test coverage gaps, mechanical refactors, small well-described bugs — and out of its depth on anything underspecified or architecturally novel. Sessions consumed real CI minutes and demanded real review attention. The teams that got value immediately were the ones that already had strong test suites and clear issue hygiene; the agent was a multiplier on existing engineering discipline, not a substitute for it. That sentence became something of a course motto.

Looking back from June 2026

The repo-as-habitat bet looks sound a year on: assigning tasks to agents through the issue tracker stopped feeling strange with remarkable speed, and the editor, terminal, and repository settled into complementary roles rather than a winner-take-all fight — most serious teams now use some combination of all three. The durable change wasn't any one product; it was that "review the agent's PR" became a routine line in the engineering day. May 2025 is when delegation moved from preview to job description.

We cover today's full Copilot stack — completion, agent mode, and the coding agent — in GitHub Copilot 2026, and Agentic Coding with GitHub Copilot focuses on the delegation workflow itself: issue specification, agent supervision, and the review discipline that makes repo-resident agents safe to scale.