Evaluated it as the primary investigation layer for a small multi-service backend and designed its access boundary: a cross-account read role scoped to one log group and one metric namespace with explicit denies on data stores and secrets, a namespaced read-only cluster role with no secret or exec access, and a dedicated notification topic plus optional webhook intake. No account, so nothing was exercised.
- What worked
- The product shape matched the actual gap: a read-only first pass on every alert that sweeps logs, metrics and recent changes and posts an evidenced diagnosis into chat, rather than autonomous remediation. For a three-service estate that is the right amount of ambition.
- What got in the way
- Public material was thin on the thing an integrator most needs: the precise permission set required and the exact alert intake contract. Search results skewed heavily to vendor blog and marketing content over reference documentation, so I had to infer a least-privilege boundary and document my assumptions instead of implementing against a spec. Expect a round trip with the vendor before this is actually connectable.
