Over-Apology Loop

Issue: Agent apologizes repeatedly without solving.

Frequency: Rare

Symptoms

• Repeated apology phrases; no resolution step.
• Conversation turn count grows for several exchanges with no new tool call, retrieved information, or concrete next step introduced.
• User explicitly calls out the loop (“stop apologizing and just fix it”) without the agent’s behavior changing on the next turn.

Root Cause The prompt emphasizes empathetic, apologetic language without pairing it to any mandatory resolution or escalation step, so apology becomes a free action the model can reach for whenever it lacks something concrete to offer — often because the actual resolution path (a tool call or policy lookup) is stuck or erroring and the model falls back to apologizing as filler rather than surfacing that failure. With no detector checking whether a turn actually introduced new progress and no cap on consecutive apology turns, nothing forces a transition to a real action or a human handoff, so the pattern can repeat indefinitely even after the user explicitly names it.

Example

User: "My package still hasn't arrived and it's been 2 weeks."
Agent: "I'm so sorry for the inconvenience this has caused you."
User: "Okay, but what are you going to do about it?"
Agent: "I sincerely apologize for the delay and any frustration this has caused."
User: "You keep apologizing. Can you just reship it or refund me?"
Agent: "I understand your frustration, and I apologize again for this experience."

Contributing Factors

• Prompt emphasizes empathetic/apologetic language without pairing it with a mandatory resolution or escalation step.
• No no-progress-turn detector, so the model can emit an apology-only response repeatedly without a validation check blocking it.
• Underlying resolution path (tool call, policy lookup) is genuinely stuck or erroring, and the model falls back to apology as filler instead of surfacing the failure and escalating.
• No cap on consecutive apology turns, so nothing forces a transition to action or human handoff.

Eval Recipes

Test Cases

Metrics

Mitigation Strategies

Prevention

1. Resolution-first response ordering: enforce a response template requiring a concrete next action or resolution attempt before, or in place of, any apology language, since the failure pattern is apology substituting for problem-solving rather than accompanying it. Trade-off: can feel abrupt or less empathetic in genuinely distressing situations where acknowledgment matters before action.
2. Apology budget per conversation: cap apology phrases to at most one per conversation (or one per newly introduced problem), forcing subsequent turns to contain only resolution content, since repeated apologizing is a low-effort filler the model defaults to when it lacks a concrete next step. Trade-off: a hard cap risks the agent sounding cold if a genuinely new issue arises later that needs acknowledgment.
3. No-progress-turn detector in generation: before finalizing a response, check whether it contains a new tool call, new information, or a concrete next step; if not, block sending an apology-only response and force a fallback (retry solve path or escalate). Trade-off: adds a validation step to every turn and can produce escalations that could have been avoided with one more apology-cushioned turn.

Detection & Response

1. Consecutive-apology-turn scanning: scan transcripts for 2+ consecutive agent turns containing apology language without an intervening tool call or new resolution content, the direct signature of this failure. Response: flag for review and auto-escalate the live conversation to a human if detected in real time.
2. No-progress conversation length correlation: track conversations where turn count grows but no new tool calls/information are introduced, correlated with apology-phrase density. Response: surface these as candidates for resolution-first template enforcement.
3. CSAT free-text mining for “not helpful”/“just apologizing” language: mine post-conversation CSAT comments for language indicating the user noticed empty apologizing. Response: pull the matching transcript into the eval set as a labeled failure example.

Architecture Patterns

1. Resolution state machine with apology as a bounded sub-state: model the conversation as a state machine where “apologize” is a transient sub-state that must transition to “attempt resolution” or “escalate” within one turn, structurally preventing indefinite apology loops.
2. Real-time loop-breaker escalation gate: a runtime guard that detects N consecutive no-progress turns and forces escalation to a human regardless of what the model would otherwise generate, acting as a backstop independent of prompt compliance.
3. Template-constrained response composer: separate “empathy” and “action” as distinct, independently required slots in a structured response composer, so a response literally cannot be emitted with an empathy slot filled and an action slot empty.

Metrics

1. consecutive_apology_turns_rate: Target: <1% of conversations have 2+ consecutive no-progress apology turns; Alert on >3%
2. no_progress_turn_rate: Target: <5% of all agent turns; Alert on >10%
3. apology_without_resolution_csat_mentions: Target: <2% of negative CSAT comments; Alert on >5%
4. loop_breaker_trigger_rate: Target: <1% of conversations trigger the escalation backstop; Alert on >3% (indicates upstream prevention is failing)

Alerts

1. Apology Loop Detected Live (P2): Condition - real-time detector finds 2+ consecutive no-progress apology turns in an active conversation. Action: auto-escalate to a human agent, attach a loop-detection flag to the handoff.
2. No-Progress Turn Rate Spike (P2): Condition - no_progress_turn_rate exceeds 10% over 24h. Action: check for a recent prompt regression removing resolution-first instructions.
3. Loop-Breaker Overuse (P3): Condition - loop_breaker_trigger_rate exceeds 3% weekly. Action: review the prevention layer (resolution-first template) for gaps, since the backstop is firing too often.

Production Signals

Key Metrics

Alerts

References

• NIST-GenAI-Profile
• Note: Generative AI risks including confabulation, data privacy, information integrity, human-AI configuration, security, value chain.