Grit Agent

The Grit Agent is a powerful agent capable of making large scale changes to codebases. It is designed to execute tasks autonomously and continuously run in the background to improve your codebase.

The Grit Agent follows a flow with three main phases: Planning, Generation, and Refinement.

In the planning phase, Grit performs an initial analysis of the codebase to find the right code to modify and to generate a high level plan on how to modify it.

This process combines static analyzers (based on the GritQL query language) and LLMs to generate an index of the codebase that covers information like code dependencies, core functionality of each file, and structural facts about class hierarchies.

After the plan phase is complete, Grit has a rough plan for how to fix the code. This plan is then passed to the generation phase.

In the generation phase, Grit uses the plan to rewrite the code in a controlled and secure way. It checks out the codebase into a secure sandbox and applies a series of transforms—some static/deterministc, and some powered by AI—that were defined during the planning phase.

When necessary, Grit will also query external tools during the generation phase. For example, documentation about libraries will be automatically fetched from the internet when it is needed.

During the refinement phase, Grit is responsible for validating changes and refining them to a level of mergeability.

Refinement is an iterative process where Grit runs external validation tools from your CI pipeline, including type checkers and unit tests, to ensure that the changes are of high quality. Grit will iterate on the changes until all validation checks pass—sometimes requiring hundreds of iterations.

During this phase, Grit will also run a self-review of the changes before they are submitted for a final code review by developers. This pre-review is guided by a set of default rules, but can also be customized by the user through autoreview rules.

Once the refinement phase is complete, Grit submits pull requests for the changes to be reviewed by a human where additional feedback can be provided.