AWS Machine Learning vs LiveKit Agents

1.7.0 is the first release in this window that is not provider breadth or failure-path repair. It adds PII redaction to Agent Observability — semantic redaction of detected entities from chat history and audio recordings, with sensitive fields filtered out of logs and traces while diagnostic context survives — and renames trace attributes and log fields to support it, which breaks third-party observability queries on upgrade. The same release adds expressive mode, where a voice agent's prosody and emotion are set by emotion tags the model generates from conversation context rather than by configuration. Underneath, the usual run of turn-taking fixes continues: tool events emitted after interruption, adaptive interruption preserved across tool calls, transcripts kept when TTS returns no word timings.

The train has been a breadth-plus-correctness operation — add speech and avatar vendors, then fix the ways conversations go wrong, with endpointing recurring constantly. 1.7.0 points somewhere else: at what a voice agent is allowed to record and how it is allowed to sound. Both are properties of the platform rather than of a plugin, and the trace-attribute rename shows the observability layer being treated as a product surface with its own contract. Cadence stays roughly weekly with a largely external contributor list.

Redaction policy will need to become configurable — which entity classes, retained or dropped at capture — since a single semantic default will not satisfy both debugging and compliance. Expect expressive mode to grow explicit overrides once developers find the model choosing the wrong tone.