LiveKit Cloud API lets coding agents read voice-agent calls: one run over 4,321 sessions returned 8 ranked issues
Synopsis
LiveKit released the LiveKit Cloud API, which lets a coding agent read the transcripts, logs, traces and recordings of a voice agent's calls and cluster the problems into an impact-ranked list. Previously such diagnosis meant reading individual calls by hand. On LiveKit's own docs assistant, the coding agent read 4,321 sessions from the past week and returned 8 ranked issues after about 30 minutes, finding that 28% of sessions were automated traffic and that slow turns came from chained tool calls.
Interpretation
A voice agent's call records can now be read and summarized directly by a coding agent, so operators get an impact-ranked issue table instead of listening through calls one at a time. Previously, locating a voice agent's problems meant reading the transcripts and logs of individual calls and judging by hand which problems affected the most calls. On LiveKit's own docs assistant, the coding agent read 4,321 sessions from the past week and returned 8 ranked issues after about 30 minutes of work, reporting that 28% of sessions were automated traffic and that slow turns came from chained tool calls.
The approach has the coding agent pull session data through a command-line tool and cluster what it finds into categories such as slow turns, tool errors, interruptions, misunderstandings, knowledge gaps and requests for a human. The LiveKit Cloud API exposes read access to transcripts, logs, traces and recordings, and the accompanying LiveKit CLI 2.19.4 or later plus a skills install gives the coding agent those commands. The sample output attributes 61 of 412 calls to order-status lookups averaging 4 seconds, 38 to the agent reading 12-digit confirmation numbers in one breath, and 27 to a missing return policy in the agent's instructions.
Perspective
The demonstration covers a voice agent on LiveKit Cloud that has taken calls with Agent Observability turned on, operated by someone with the project admin role and LiveKit CLI 2.19.4 or later. For agents with many calls, the text suggests starting from a sample such as the last 100 sessions or the last 24 hours, because the coding agent reads every transcript it pulls.
The reported figures come from a single run on LiveKit's own docs assistant, and the text does not report performance on other agents or larger samples. The sample issue table is based on 412 calls, a different set from the 4,321-session demonstration.
