lever: Feedback Loops for Coding Agents
GitHub has introduced Lever, a new Copilot plugin that creates feedback loops for coding agents, allowing developers to automate repetitive tasks and improve code quality through measurable metrics.
The plugin introduces a system where developers can describe problems in natural language, and Lever identifies appropriate workflows to address them. For example, typing '/lever loop our test suite feels slow' generates a Loop Card with metrics, measurement methods, noise levels, and success criteria without making any code changes.
Lever addresses a common challenge in AI-assisted development: the difficulty of setting up effective feedback loops. By asking developers what to measure, how noisy the metric is, and what constitutes a real win, the plugin automates the process of creating measurable improvements for AI agents to work toward.
How it works
The plugin incorporates various commands that enable different workflows. '/lever arms' runs multiple agents in parallel with different ideas against a control, while '/loop' focuses on making something measurably better. '/automate' turns repetitive chores into paused Copilot Automations that can be activated when needed.
Developers can customize Lever's behavior through several dials. Autonomy ranges from L0 (only proposes changes) to L5 (lands, releases and watches), determining how much control the agent has. Spend settings control model tier, parallel lanes, review depth, and execution frequency.
Implementation and customization
Lever can be installed through the Copilot marketplace or manually by copying its skills, documentation, templates, and configuration files to a repository. The first run prompts developers with three questions to set up their environment, with team configurations managed through .github/lever.yml files.
These skills fit how we work. Yours will be different. /lever mirror reads your own session history and shows you what you keep repeating, and /lever steer turns those repeated corrections into rules that stick.
Future development
The project is currently in early development, with GitHub encouraging community contributions through issues and pull requests. The plugin is released under an MIT license, allowing developers to fork and customize it according to their specific needs and workflows.
Automations within Lever are designed to be safe, starting as paused configurations that dry-run on past events before activation. This approach ensures that automated processes are thoroughly tested before they begin affecting code repositories.
Related
