GitHub Copilot

Copilot started as autocomplete and is now four things stapled together: inline completions, a chat sidebar, a cloud coding agent that opens pull requests on its own branch and a CLI. Custom agents are markdown files, agent profiles that declare a prompt, a tool list and a set of MCP servers, which means a team can encode its review conventions once and have every agent run inherit them. Completions stay free and unlimited on paid plans. Agent runs and frontier model calls are what draw down credits.

The reason engineering leads keep it on the shortlist has little to do with model quality. Microsoft indemnifies customers against copyright claims on suggested code. Admins can disable specific models per organisation, enforce public-code filtering and pull audit logs. It runs in VS Code, Visual Studio, JetBrains IDEs, Neovim and Xcode, so you are not forcing a whole team into one editor. For a 200-person org already on GitHub Enterprise, procurement is a checkbox rather than a six-week security review.

Where it loses is autonomy. On work that needs several files changed in sequence with test feedback in the loop, Cursor's agent and Anthropic's terminal agent both get further before a human has to step in. Copilot's agent is good at the well-scoped ticket: add the test, fix the lint rule, bump the dependency and explain the diff. Treat it as the org-wide floor and let the teams doing hard refactors expense something sharper.