Description
The agent sometimes responds with "I'll look into that for you" or "I can definitely do some research on that" and then stops (stop=end_turn) without actually calling any tools. The user is left waiting for research that never happens.
Reproduction
/start to reset session- Ask a factual question (e.g. "What's 450k vnd to usd") — agent correctly uses
web_search and responds
- Follow up with a research request: "What about can you tell me about oil crisis if it's coming to Vietnam too? Do some research on it"
- Agent responds: "I can definitely do some research on that for you. I'll look into the current oil crisis and whether it's projected to affect Vietnam."
- No tool calls are made. The agent just stops.
Audit log evidence
Only one tool call was logged for the entire session after the research request — zero web_search, web_fetch, or delegate_task calls:
210|2026-03-19T11:22:49Z|web_search|{"query":"450000 VND to USD"}|... (this is from the first question)
Nothing after that for the oil crisis question.
Expected behavior
The agent should immediately call web_search (or delegate_task for deeper research) instead of narrating its plan. The system prompt already says: "ALWAYS use tools to perform actions. Never say you will do something without calling the tool." — but the agent ignored this.
Context
The system prompt has the right instruction, so this is likely an LLM-level compliance issue rather than a missing prompt. May need stronger reinforcement or a retry mechanism when the agent produces a text-only response to an actionable request.
Description
The agent sometimes responds with "I'll look into that for you" or "I can definitely do some research on that" and then stops (
stop=end_turn) without actually calling any tools. The user is left waiting for research that never happens.Reproduction
/startto reset sessionweb_searchand respondsAudit log evidence
Only one tool call was logged for the entire session after the research request — zero
web_search,web_fetch, ordelegate_taskcalls:Nothing after that for the oil crisis question.
Expected behavior
The agent should immediately call
web_search(ordelegate_taskfor deeper research) instead of narrating its plan. The system prompt already says: "ALWAYS use tools to perform actions. Never say you will do something without calling the tool." — but the agent ignored this.Context
The system prompt has the right instruction, so this is likely an LLM-level compliance issue rather than a missing prompt. May need stronger reinforcement or a retry mechanism when the agent produces a text-only response to an actionable request.