Engineering AI tools fall on a spectrum of autonomy.
1. Inline completion
Suggests the next lines as you type. Low risk, low setup, modest gains; the developer sees and accepts every suggestion.
2. Chat and edit in the IDE
The developer asks for a function, a refactor or an explanation; the tool proposes multi-file edits. The developer still drives and reviews each change.
3. Agent mode in the IDE or terminal
The tool plans, edits several files, runs commands and tests, and iterates, with the developer watching and approving steps. Gains on larger tasks can be big; so is token use.
4. Background (asynchronous) coding agents
You assign a task — often an issue — and the agent works on its own in a sandboxed environment, then opens a pull request for human review. GitHub made its Copilot coding agent generally available in September 2025: you delegate a task, and Copilot opens a draft pull request and works in the background in its own development environment powered by GitHub Actions, then requests review.
What changes as autonomy rises
- Work shifts from writing to specifying and reviewing. A good issue description becomes a key input.
- Review load grows. Agents can open many PRs; humans still approve them.
- Permissions matter more. A background agent needs its own identity, scoped access and an environment it can't escape.
- Costs become usage-based. Agents consume far more tokens than completion.
Fit to task
Background agents do best on well-specified, verifiable tasks: test coverage, small bug fixes, dependency upgrades, documentation, migrations with clear rules. They struggle with ambiguous product work and changes that need architectural judgment.