Repository architecture
Git-native AI software production treats immutable commits and a customer-owned repository as the durable code record. AI agents work in isolated candidate state, evidence refers to exact SHAs, and one integration authority advances the canonical history. The software remains portable beyond the orchestration platform.
Agent chats and vendor task logs are useful context, but they are not a portable source tree. Git preserves the code, history and exact identities conventional engineering tools already understand.
Nexus layers orchestration state around Git rather than replacing it.
A worker can create a technically valid commit without authority to publish it as the shared product. Review and integration are separate transitions.
This distinction keeps parallel AI agents from pushing competing interpretations directly into the canonical branch.
Stale baselines, merge conflicts and deleted workspaces are expected distributed-system events. Durable state records which candidate existed and why it was accepted or rejected.
A multi-agent Git workflow should expose conflicts for remediation, not hide them behind forced pushes.
Nexus launch access
Register your interest now and we will let you know when the autonomous AI software engineering control plane is ready.
