LiveKit End-to-End Concurrent Call Benchmarks

Summary

Capacity findings for single-bot LiveKit (pipeline_flow_agent / Expertflow-agent) on VM 194.146.13.204 (12 vCPU, ~47 GB RAM). Stack: LiveKit server + LiveKit SIP + Redis colocated; inbound DN 10005; transfer to FusionPBX 99900 → EFCX.

Date: August 2026

Distinct caller IDs are required for concurrent E2E (same CID collapses onto one desk). ElevenLabs TTS concurrency hard-cap = 15.

Bot-only (no human transfer)

N

Result

Peak load

Peak SIP CPU

Peak LiveKit CPU

Notes

5

PASS

0.71

~19%

~15%

Short hold

10

PASS

12.18

~84%

~46%

Sustained ~3–4 min

15

PASS

16.93

~81%

~53%

TTS gate; N>15 not valid

Ops target (this VM): ~12 concurrent bot-only.

E2E (bot → human via 99900)

N

Result

Peak load

Peak SIP

Peak LiveKit

Notes

1

PASS

light

Full transfer + CCM

3

PASS

light

Distinct CIDs

5

PASS

8.99

~54%

~28%

Voice-confirmed

10

PASS

16.14

~88%

~42%

10/10 answered

Capacity takeaway

Mode

Proven / guidance

Bot-only

5 / 10 / 15 proven; ops ~12; prefer SIP scale-out or ≥16 vCPU before >12

E2E

1 / 3 / 5 proven; 10 staffing-limited; watch LiveKit SIP CPU first

Cross-call drops during harness hangup (sched_hangup / hupall) are normal for the test harness, not a product multi-room bug.

Repo: agent/docs/TEST_FINDINGS.md.