“ So, what could go wrong?,” Jason asks sarcastically as these OpenClaw instances with such deep access crawling about the ecosystem is a disaster waiting to happen.
Another aspect that goes hand in hand with AI visibility is protecting the system itself against data poisoning, prompt injection, prompt fuzzing and the many threats that enterprise LLMs are susceptible to.
Micromanaging agentic AI: observable and accountable The observability question falls right into Mayank Agarwal’ s court. Agarwal describes the non-deterministic nature of these agents as a strength.
“ They are able to explore, reason from first principles and solve problems that may not have been previously solved because this is a new kind of problem that just got introduced into production,” he explains.
“ I would say that because these agents are basically doing a lot of exploration and creating these decision traces, you have to capture the entire decision traces and explain to the end user what was the proposal or what was the action that was taken by the agent.”
He says it is“ very important that every step of reasoning is backed by a citation to an underlying system of regard that displays the truth that the agent used to come to its full chain of reasoning”.
112 September 2026