Published 25 October 2026 · dotNewLABS
Every vendor is selling AI agents, and good tools now exist to watch them. They do a lot. Here is what they do, and four things we see them leave to people.
A safety layer sits between your agents and your data and does three jobs: see every agent and keep a live list, check what each one can reach, and stop what is not allowed, as it happens.
That split is sensible. A rule that blocks an action should behave the same way every time, and be easy to explain to an auditor.
1
Setup and tuningSomeone has to connect each agent, write the rules for this company, and fix what the tool flags. The software does not know your policies.
2
Testing on your own casesSafety tools check for attacks and misuse. They do not tell you whether the agent gave the right answer to your customer's question.
3
Proof a leader can signA board or an auditor wants a plain report of what was tested and what happened, not a security console to read.
4
Safe doors into older systemsAn agent is only as safe as what it can reach. Home-grown apps and old databases often need narrow, fixed ways in before an agent should touch them.
Software companies build products that scale. Setting up around one company's policies, testing against its own past cases and wiring safe access into its older systems is work for people, and it does not scale the same way. Larger firms can pay a big consultancy to do it. Smaller ones often have the budget for the software and nobody to run it.