One operator, a fleet of agents

The human operator in this system acts as the gatekeeper, rather than a worker. The primary goal is for the operator to decide which changes should be made, rather than typing every change themselves. To achieve this, every unit of work that an agent can handle must be approved by the human operator before an agent begins working on it.

This ensures that the agent's actions are controlled and safe. The fleet of autonomous coding agents is powerful because it can handle multiple tasks in parallel, rather than being limited to a single agent. However, the agents are not completely autonomous, as they must wait for human approval before starting any work. The approval process is streamlined, with the human operator able to approve multiple changes from a single dashboard.

This approval queue allows for quick decision-making, but does not automatically approve changes without human input. The system also includes a deploy block step, which requires manual approval before irreversible changes are made to the live server. In addition to being controlled by the human operator, the fleet of agents also includes built-in safeguards to prevent conflicts.

Each agent works on its own copy of the repository, ensuring that changes do not interfere with each other. The system then returns the results of each agent's work as reviewable diffs, allowing for easy comparison and review. The system also includes a mechanism for reporting back to the operator, allowing them to keep track of progress and any issues that arise.

Overall, this system combines the efficiency of a fleet of autonomous agents with the safety and oversight of a human operator, creating a powerful and reliable tool for managing a portfolio of sites.