LLMs, GenAI, RAG, agents, evaluation · insight-1553

Detect repeated identical tool calls and break the loop before the budget runs out

Situation
An agent searches, gets an unhelpful result, and issues the same call with the same arguments again and again.
Action
Keep a per-run record of tool name plus normalised arguments, and when the same call repeats a small number of times without new information, stop the loop and either change strategy, ask the user or escalate with the trace.
Reason
A hard budget bounds the damage but only after the waste; a repeat check recognises the stuck pattern in two or three steps and gives a more useful failure.
Boundary
Any agent loop with search, lookup or retry-style tools.

Symptom or risk

The run burns its whole step and token budget producing nothing, and the user waits until the cap fires.

Polling a job status legitimately repeats a call; give such tools a separate, timed allowance.

Argument normalisation takes care, and near-duplicate calls with trivial changes can slip past.

Treating the model's own wording changes as new progress when the retrieved data is the same.