autoresearch

Autoresearch automates iterative codebase improvements by defining a single mechanical metric and executing one atomic change per cycle. It analyzes context, modifies code, verifies impact, and uses Git history to decide whether to keep or revert changes.

Git history serves as memory to guide autonomous code refactoring decisions. The system learns from past iterations to determine whether to keep, discard, or revert modifications, ensuring improvements are auditable and repeatable across development cycles.

Constraint-driven planning translates plain-language goals into executable iteration loops. It bounds automated modifications to atomic changes per cycle, supporting software engineering, security, documentation, and shipping tasks with verifiable and auditable progress.

Bounded iterations for CI/CD gates and overnight experiments are configured by invoking the inline command with a defined goal and iteration limit. Alternatively, launching the planner generates a full configuration to automate and constrain the execution scope.

Autoresearch supports domain-adaptive workflows beyond software engineering, including security, documentation, and shipping use cases. It adapts its constraint-driven iteration loops to plan scope, execute modifications, and verify impact across these varied domains.

Autonomous iteration loops for code modifications should be avoided when a task cannot be distilled into a single mechanical success metric. Without an atomic, verifiable change per cycle, the system cannot effectively verify impact or log meaningful results.