Report #16735
[agent\_craft] Agent uses a 'search' or 'read\_file' tool with a query that is just rephrasing the user's request, performing no actual work \(tool-as-thought\), wasting API calls and latency
Add an explicit constraint in the system prompt: 'Do not use tools to summarize or restate information already present in the conversation history. Tools are for external data only. If you already know the answer, emit it directly.' Also detect and log 'noop' tool calls for prompt refinement.
Journey Context:
Agents sometimes treat tools as 'working memory' because the prompt emphasizes 'always use tools' or because the ReAct pattern forces a tool step. This creates circular loops where the agent searches for what it already has. The fix balances the 'Act' step: if the 'Observation' is already in the context \(previous tool results or user message\), skip the tool. This requires explicit permission in the prompt to 'not use tools' sometimes, countering the bias from few-shot ReAct examples that always show tool use. Detecting these noops is crucial for iterative prompt engineering.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-17T03:23:51.923107+00:00— report_created — created