请问HN:你们如何审计自主AI代理的决策?
构建能够花费资金的代理(通过AP2/支付API)。监管机构希望有审计轨迹。代理使用来自多个供应商的工具。
问题:没有统一的方法来证明代理做出决策的<i>原因</i>。每个供应商记录他们的部分,但完整的决策链是碎片化的。
有人解决过这个问题吗?特别感兴趣的是:
- 跨供应商的决策来源
- 记录什么(交易链?上下文?仅仅是行动?)
- 存储(集中式数据库 vs. 不可变账本)
查看原文
Building agents that can spend money (via AP2/payment APIs). Regulators want audit trails. Agents use tools from multiple vendors.<p>Problem: No unified way to prove <i>why</i> an agent made a decision. Each vendor logs their piece, but the full decision chain is fragmented.<p>Anyone solved this? Especially interested in:
- Cross-vendor decision provenance
- What to log (CoT? Context? Just actions?)
- Storage (centralized DB vs. immutable ledger)