Ankole team analyzes four reasons multi-user agent sharing goes out of control, shares open-source Business Agent architecture design
Agent多人共用为何会失控?来自Ankole的Business Agent架构设计分享
The Ankole team shared the architecture design of its open-source Business Agent system, saying that an agent that performs stably on its own fails in four ways once it enters a multi-person organization: cognitive divergence, context window leakage, state fragmentation, and the compute tax and compounding error of polling; for example, with 99% accuracy per judgment and a check every 5 minutes, 288 rounds a day, the probability of at least one error within a day is about 94.5%. Its Business Harness hands routine flows to scripts written by the model, with the model handling only exceptions; before each wake-up it takes the intersection of the permissions of everyone present, and information that any one person is not authorized to view does not enter the context; organizational memory shares a relationship directory containing no secrets and checks source document permissions in real time. The system orchestrates concurrency with the Actor model, and irreversible operations are first recorded in a persistent action ledger; the team says it has used Lean 4 to formally verify the core permission state machine, and the code is open-sourced at https://github.com/AgentBull/ankole。