Why do agent guardrails and permission mapping fall apart once agents call other agents?

Straightforward IAM gets messy fast once agents start orchestrating other agents or calling tools dynamically based on context. A single top-level permission grant doesn't tell you much about what actually happens three hops down the chain. We're trying to build a real permission map, not just "this agent has API access to X" but the actual reachable graph of tools and data across delegated calls. Doing it manually stopped scaling past a handful of agents. The mixed environment makes it worse. Homegrown agents, SaaS agents, cloud-native agents, each exposes permissions differently, and stitching that into one coherent map is its own project. Is anyone doing this with real tooling instead of docs that go stale in a week? Specifically interested in approaches that don't require rewriting every agent to emit permission telemetry, since that's not realistic across a mixed stack.