Sentri pitches safer enterprise agents

This is more compelling as a safety-and-systems paper than as a generic “agent” story. If the author can back the claims with rigorous ablations and realistic failure analysis, Sentri could land well in VLDB/SIGMOD/MLSys; if the contribution stays mostly conceptual, it reads closer to an AI-safety workshop piece.

• –The strongest angle is production containment: automation only matters here if the system reliably prevents destructive actions without freezing too many safe ones.
• –The three-layer design is sensible defense in depth, but the paper will need to show each layer catches distinct failure modes, not just stacked redundancy.
• –“Production-safe” should be measured with policy violations prevented, safe actions incorrectly blocked, end-to-end task success under constraints, and adversarial red-team prompts.
• –Deep evaluation in one domain will likely be more credible than shallow coverage across DB, cloud, and DevOps unless the same safety harness is reused consistently.
• –Baselines should include deterministic runbooks/rule engines, naive agents, and constrained-orchestrated systems with approval gates, not just an unconstrained LLM.