Local governed software orchestration
Can a development loop preserve intent while it changes the route?
Continuous Loop attaches to a Git repository and coordinates an Architect, Coder, Auditor, and Drift agent around a human-owned North Star. It persists directives, evidence, and recovery state so implementation is followed by correctness and alignment checks before another bounded decision.
Blog entries on recent progress from the researchers
No entries yet. New project notes will appear here.
The Architect decides the next bounded directive, the Coder implements only that directive, the Auditor judges whether the implementation is correct, and the Drift agent checks whether the cumulative project still follows the human-owned goal. Audit failures force repair; drift failures force realignment.
Continuous Loop keeps process-loop state separate from the target project’s workflow state. It uses a persistent .agent workspace, structured directives, Git evidence, and provider-specific execution paths rather than treating an activity panel or a single successful subprocess as durable proof.
The documented baseline is explicit about caveats: Auditor and Drift are sequential, pause does not interrupt an active agent, dirty Atlas mirrors are not clobbered, and crash recovery can rerun a coder in the same cycle directory. Those semantics are part of the system’s contract and must remain visible when the loop evolves.
Continuous Loop is an orchestration and evidence system. It does not itself implement a target product or guarantee that an AI-generated change is correct.
Read the source record in Notion ↗