A human organization can take on large undertakings because roles are divided, rules exist for handing work over, and someone oversees the whole. AI is no different. Rather than making a single AI larger, we divide the work among AIs with different roles and coordinate them centrally, which greatly widens the volume and variety of work that can be handled.
We create agents for each role, such as collecting, judging, drafting, and checking, define the format of each handoff, and build in failure detection, automatic retries, and notifications to people. A human approval is placed at every step that requires judgment, and only approved items are executed.
Our AI bidding system rereads public information four times a day, judges approx. 100,000 items, and keeps tracking 1,108 cases. The multi-agent that oversees the operation of our entire business compiles the status of all businesses with seven AI agents.