ElevenLabs · Termination Insights
2,261 calls · deterministic (no LLM)
Note — excluded data: Excluded from these insights: (1) 49 historical 'exceeded maximum duration' cutoffs (Kazumi, Apr 15–17 2026, ~3-min cap since raised to 60 min) — a config artifact; (2) 'exceeds your quota limit' rejections — account-level credit/usage-cap refusals that bounce calls ~2s at connect (infra, not call-termination behavior). Both removed to avoid polluting insights. See exclusions.py.
32
clear (>15s silent)
0
never spoke at all
138
borderline (8–15s)
2
premature (<8s)
Measured signal: seconds the user had been quiet (call end − last user utterance) when the agent ended for silence. Most cluster in the 8–15s borderline band — right at the agent's re-prompt threshold — so a definitive true/false-positive verdict needs the LLM pass, not this timing heuristic.
agent said: “user not responding”
User's last words at 33s — only 4s of silence before the agent ended.
agent said: “User has gone quiet and is not responding.”
User's last words at 87s — only 4s of silence before the agent ended.
| Agent | Calls | Agent-ended | Abnormal drop | Silence TO | Normal |
|---|---|---|---|---|---|
| PureOceanEyes | 22 | 32% | 9% | 0% | 55% |
| Cowgirl J | 18 | 28% | 6% | 0% | 61% |
| Rachel Jane | 16 | 25% | 0% | 0% | 75% |
| Nicole Thibeault | 16 | 25% | 6% | 0% | 62% |
| Tayuh Lynn | 23 | 22% | 13% | 0% | 61% |
| Sophie Dee | 823 | 22% | 2% | 6% | 67% |
| Desirae Fraser | 99 | 18% | 5% | 1% | 74% |
| Mrs Poindexter | 29 | 17% | 7% | 0% | 76% |
| Rebecca J | 100 | 16% | 6% | 1% | 70% |
| Ana Nello | 27 | 15% | 0% | 0% | 81% |
| Beatriz | 7 | 14% | 0% | 0% | 86% |
| Lena The Plug | 279 | 13% | 4% | 0% | 81% |
| CeCe Lynn | 31 | 13% | 6% | 0% | 68% |
| Kiki D'Aire | 49 | 12% | 6% | 0% | 80% |
| Kazumi | 184 | 11% | 2% | 1% | 76% |
| Savannah Bond | 40 | 10% | 8% | 0% | 80% |
| Princess Frenxh | 50 | 10% | 6% | 0% | 78% |
| Alexis Mucci | 124 | 9% | 3% | 1% | 80% |
| Brody Jean | 23 | 9% | 4% | 4% | 83% |
| Ashley Belle | 22 | 9% | 5% | 0% | 82% |
| Miss Lexa | 61 | 8% | 2% | 2% | 87% |
| Ellie Meyers | 25 | 8% | 0% | 4% | 84% |
| Nova Frog | 33 | 6% | 15% | 0% | 76% |
| Nicole | 22 | 5% | 0% | 0% | 91% |
| Janna Moller | 37 | 5% | 5% | 5% | 76% |
| Amazon Victoria | 19 | 5% | 11% | 0% | 84% |
| Anastasia | 19 | 5% | 5% | 5% | 79% |
| Reese | 43 | 5% | 5% | 2% | 84% |
| Lisa Daniels | 8 | 0% | 12% | 0% | 88% |
| Kylie Kennedy | 2 | 0% | 0% | 0% | 100% |
| Barbara Silva | 10 | 0% | 0% | 0% | 90% |
Red agent-ended ≥25% / amber abnormal-drop ≥10% flag outliers worth a look.
<5s: 88 · <10s: 302 (13% of calls)
<10s by reason: User disconnected 265, User left 14, Connection dropped 11, Non retryable error during text to speech 9
Impact ●●● high → ○○ low · Effort ●●● high → ○○ low. P0 = do now (high impact, prompt-only).
| Pri | Step | Impact | Effort |
|---|---|---|---|
| P0 | Silence → re-engage (prompt + end_call description edit) Biggest single lever — ~48% of agent hang-ups. Prompt-only change. |
●●● | ●○○ |
| P0 | Block explicitness-triggered ends; redirect to CTA Highest-volume false-positive cluster; already against the agents' own rule. |
●●● | ●○○ |
| P1 | Define 'Stage 4' as a re-engage loop (remove arc-complete close) Removes the improvised-close behavior at the prompt level. |
●●○ | ●○○ |
| P1 | Extend `no_premature_end` eval to silence/arc/disengagement Turns the fix into a measurable metric per agent. |
●●○ | ●○○ |
| P1 | Run the LLM policy review (add ANTHROPIC_API_KEY) Precise per-call verdict + per-agent offender ranking; tightens the needs-human set. |
●●○ | ●○○ |
| P2 | Human review the ~54 needs-human cases; codify gray-area policy Taboo content, fantasy-vs-real non-consent, user-initiated exits. Needs reviewer time. |
●●○ | ●●○ |
| P2 | Ops fix: account quota/credit exhaustion + TTS errors Calls failing at connect waste first impressions. Separate team. |
●●○ | ●●○ |
| P3 | Instrument user identity for new/returning analytics Requires eng + data work (call_id → backend join or anon-id passthrough). |
●○○ | ●●● |