8 Comments
User's avatar
Marcus Kim's avatar

The periodic registry refresh makes emergency disablement an interesting failure case to test. I'd disable a previously approved tool while one gateway replica still has the old in-memory mapping, then check whether the next call is rejected before reaching the downstream service. Starting newly discovered tools disabled is a good admission gate, but revocation needs a separate guarantee. Is there a fast invalidation path for that case, or is the refresh interval the intended upper bound on disablement taking effect?

Aarti Sharma's avatar

Your breakdown of Uber's AI agent gateway raises interesting questions about secure API integrations; do you also work with fintech teams building payment or financial infrastructure?

amar's avatar

Tool-level authorization is the part most internal gateways skip, so it's good to see it spelled out with the incident example, where reading and changing status get different policies. The gateway also seems like the natural place to handle the other direction. It redacts PII from tool responses, but those same responses can carry text someone typed into a Jira ticket, and that text goes straight into the agent's context. Does Uber treat tool responses as untrusted input at the gateway, or leave that to each agent?

Early Signal's avatar

Agree with keeping tools off until someone turns them on. I'd also start smaller than most teams do. Build a basic version of the agent first so everyone understands how it works, then add to it, instead of planning the whole orchestration up front.

Jordan Vale's avatar

The strongest design choice here is that newly discovered tools start in a disabled state until the owning team approves them. AutoCrawler can find thousands of tools, but discovery is not permission, and that distinction is what keeps a gateway from becoming a back door. My question is about the other end of that lifecycle: once a tool is approved, how fast does a revocation actually propagate? If the data plane refreshes periodically, there could be a window where a disabled tool is still callable, and for a small team that window is the whole security story.

Nan Deng's avatar

The part I would not loosen is that AutoCrawler registers newly discovered tools in a disabled state until the owning team approves a diff. I would bind that approval to the exact tool schema and arguments, not to an agent's summary of the call: a policy that allows reading incidents should not also allow changing status just because both tools live on the same virtual server. Omni MCP's discover_server and get_tool_schema split keeps thousands of definitions out of the prompt, but the data plane's periodic in-memory refresh still has to drop a tool the moment the registry disables it, or the agent keeps calling a stale allow.

Raghav Mehta's avatar

Omni MCP is the part I'd love to hear more about. Gradual discovery keeps thousands of schemas out of the context, but each new tool now costs a discover_tools and get_tool_schema round trip before the real call. Do agents at Uber tend to settle on a small set of tools per task, so the fetched schemas stay early in the context and cache well? Or do long incident investigations keep pulling in new ones mid session? Same question for Response Projection. Does the model pick the fields itself, or does the tool owner set a default projection?