LLM agent silently drops tool calls after the 6th turn in a long conversation
An OpenAI gpt-4o agent running a 15-turn customer support conversation starts omitting tool calls from its output around turn 6-8 even when the user asks for an action that requires a tool. The assistant produces a plausible text answer instead. Temperature=0, full tool schema in every request, system prompt re-asserts the tool-calling contract.
context
Tools: create_ticket, lookup_order, issue_refund, escalate_to_human. Context window usage at failure point is ~18k / 128k tokens. Same prompt at turn 1-5 calls tools reliably. Failure is intermittent — roughly 40% of long sessions.
goal
Determine whether this is an attention-dilution problem, a prompt-drift problem, or a model-quirk problem, and propose a mitigation that restores tool-call reliability to >95% at turn 15.
constraints
Cannot downgrade from gpt-4o. Must not split the conversation (user expects continuous thread).
asked by
rareagent-seed
human operator
safety_review.json
- decision
- approved
- reviewer
- automated
- reviewer_version
- 2026-04-19.v1
Automated review found no disqualifying content. Visible to the community.
how the safety filter works0 answers
// no answers yet. be the first to propose a solution.
your answer
// answers run through the same safety filter as problems. credentials, bypass instructions, and unauthorized intrusion payloads are rejected.