The two LLM-call model pre-decides every thing earlier than execution. However in the event you give the second name entry to a software or agent as a substitute of simply producing a response, then it might nonetheless re-route mid-task if it feels the plan isn’t proper. This recovers some agentic flexibility. Confidence-gated routing could possibly be one other resolution. Run the two calls first; if the arrogance is excessive, proceed utilizing it. If not, then transfer to 4 calls.
Problem 5: AI can discover the best drawback however misjudge the repair
System audits (for accuracy, latency, safety, and so on.) utilizing AI instruments are selecting up velocity. These instruments are nice at figuring out gaps, suggesting alternate options or fixes, and figuring out how a lot achieve that step would guarantee. Most significantly, these evaluations are fast, and anybody can do it. In one among our initiatives, our buyer used Claude for a similar, and the system points it recognized have been all spot-on.
The issue is these instruments will not be fully dependable. Directionally, these instruments are sometimes proper. However the fixes they recommend and the projected positive factors they present will not be at all times aligned and proper. They’ll determine the best lever solely to misjudge what’s pulling them or to make use of the fallacious metrics to calculate advantages. In our case, the software advised a latency drop by 5-6s when the achieve was solely 500 milliseconds.
