Salesforce 论文：智能体互相协作时从不主动核实对方说法
Salesforce 跑了 2500 多场智能体互相对话的实验，发现它们互相不查证还泄密，做多智能体系统的都该看看。
Salesforce 发布新论文，研究代表不同利益方的智能体之间的协作失败问题。团队用 Claude Code、Codex 和 Cursor 在隔离沙箱里跑了超过 2,500 次智能体对智能体的会话，覆盖 488 个场景，归纳出 20 种失败模式。结果显示没有任何智能体会重新拉取文档去核实同伴的事实性声明，主动联系同伴基本只在提示词要求时才发生。在一个服务纠纷案例中，一笔 24,000 美元的和解通过了全部 10 项结果检查，同时供应商方的智能体泄露了机密信息，暴露出团队整体会继承最弱智能体的错误。
New Salesforce paper finds that agents acting for different parties never verify each other's claims unprompted, so teams inherit their weakest agent's errors.
Agents increasingly coordinate with agents that represent other people or companies, each with private files, tools, and goals. Existing failure studies assume 1 orchestrator or a shared workspace.
They ran over 2,500 agent-to-agent sessions across 488 scenarios, using Claude Code, Codex, and Cursor in separate sandboxes. They catalogued 20 failure modes from the traces.
No agent re-fetched a document to test a peer's factual claim, and agents contacted peers mainly when the prompt told them to. In 1 service-level dispute, a $24,000 settlement passed all 10 outcome checks while the vendor agent leaked confidential details.
If your agent talks to outside agents, build checks into the protocol and score each agent against the party it serves, not the team result.
