What we ship, when we

Conversation-flow agents in alpha, plus eleven new Indic voices.

A meaty release this month — the long-promised structured agent runtime is now in alpha, and the Sarvam Bulbul-v3 voice pack covers eleven Indic languages out of the box.

• NewConversation-flow agents (alpha) — build multi-step structured flows alongside the single-prompt runtime.
• NewSarvam Bulbul-v3 added to the TTS provider list. Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Assamese.
• ImprovedSession viewer load time cut by 4× on workspaces with > 50k sessions.
• ImprovedOutbound call ringing now respects locale-specific DNC rules automatically.
• FixedWhatsApp message delivery delays during high-volume campaigns.
• FixedKB ingest occasionally double-counted PDF pages with embedded images.

MCP endpoint goes generally available.

After a quiet beta with a handful of design partners, the MCP endpoint is open to everyone. Connect Claude, Cursor, ChatGPT, or any MCP client and your sessions / KBs / campaigns become one prompt away.

• NewMCP endpoint at mcp.saaya.ai — read-only by default, RBAC-scoped per key.
• NewEight tools exposed: list_agents, get_agent, query_knowledge_base, list_sessions, get_session_transcript, get_session_summary, list_campaigns, get_campaign_status.
• NewBridge SDK published to npm (@saaya/mcp-bridge) for clients that prefer a local subprocess.
• ImprovedAPI rate-limit headers now match the conventions used by Stripe and Linear.
• FixedRace condition during agent rollback when a draft was promoted in the same second.

Strict-citation RAG, KB-gap analytics, and audit-log streaming.

• NewStrict-citation mode: agents refuse to answer when source confidence is below your threshold and escalate cleanly.
• NewKB-gap analytics — top unanswered questions ranked by frequency and severity.
• NewAudit-log streaming to S3, GCS, and Azure Blob.
• ImprovedKnowledge-base ingest now handles documents up to 1 GB without chunking the upload pipeline.
• ImprovedSentiment scoring switched to a fine-tuned model — accuracy up 14% on Indic-English code-mixed transcripts.
• BreakingThe deprecated `/v0/agents` namespace is removed. Migrate to `/v1/agents` (full migration guide in docs).

New campaign engine + WhatsApp-first cascades.

• NewRe-architected campaign engine — 3× more concurrent calls per worker pool, predictable retries.
• NewChannel cascades: try WhatsApp first, fall through to voice, end with email — all from one config.
• ImprovedLive ops dashboard now shows per-channel answer rate, sentiment distribution, and KB-hit ratio.
• FixedSome campaigns paused mid-run when the underlying number pool exhausted — now drains gracefully.

Saaya 2.0 — multimodal agents in one config.

The big one. One agent definition, every channel — voice, video, chat, WhatsApp. The provider matrix lets you pick the LLM, STT, TTS, and avatar per call without rewriting your prompts.

• NewMultimodal agent runtime — voice, video, chat, WhatsApp from a single config.
• NewProvider-agnostic LLM, STT, TTS, and avatar layers. Switch in seconds; no lock-in.
• NewVersioning + rollback — every publish is a versioned snapshot, restore in one call.
• ImprovedSub-second cold starts thanks to the new Redis config cache.
• Breaking`agent.voice_id` flat field is replaced by `agent.voice = { provider, voice_id }`.
• DeprecatedThe legacy single-channel "Voice Bot" object is deprecated and will be removed in 2.5.0.

Smart escalation, sentiment-aware barge-in, and 200ms turn-taking.

• NewSmart escalation — agents hand off to a human queue based on sentiment, complexity, or explicit request.
• ImprovedTurn-taking latency cut to a 200ms barge-in median across LiveKit + Twilio paths.
• FixedTool-call retries occasionally double-fired when the underlying provider returned 5xx mid-stream.