Skip to main content

Release notes

A rolling summary of platform changes, grouped by area. For the in-product story-level history with timestamps, see the Audit log.

Voice Agents​

  • Foundation (M1) — Voice Agents launched as a first-class, separate resource type. Dedicated navigation, list view, configuration screen, run history, deployment screen, and tracing. Per-second billing introduced.
  • Sync, Async, Realtime API modes — three execution paths. Sync bypasses queues; Async uses a dedicated voice priority queue; Realtime API runs through OpenAI's native multimodal API.
  • Latency telemetry — TTFS (Time to First Speech), STT/TTS wait times, queue wait, silence gap rate. P50/P90/P99 dashboards. Alert at P90 TTFS > 1200ms.
  • BYOK redesign (M2) — unified key selector per provider with SimplAI key (with cost label) + your own saved keys. Inline "Add new key" flow.
  • Voice cloning (M2) — custom voice ID input for ElevenLabs; voice preview button across all TTS providers.
  • Multilingual ASR/TTS (M2) — Deepgram Multilingual + Google as a new STT provider; TTS provider list auto-filters to multilingual-capable providers.
  • Trace fixes (M2) — STT and TTS spans now correctly timed and carry transcribed / synthesized text. Audio playback synced to spans in the run detail view.
  • Advanced STT config (M3) — keyterms (Deepgram Nova-3), endpointing, transcription option toggles (numerals, profanity filter, punctuate, smart format, filler words, interim results, diarization), fallback transcribers.
  • Advanced voice config (M3) — stability, similarity boost, style, speaker boost sliders (ElevenLabs); fallback voices.
  • Voice Agent node — new workflow node for invoking a published Voice Agent from inside a Tool.
  • Inbound calling — telephony provider connection (Karix, Twilio); phone number assignment with one-agent-per-number rule.

Tools & AI Chains​

  • Conversation mode — Tools and AI Chains unified. Every Tool has a new Conversation mode toggle. AI Chains as a standalone primitive removed; existing AI Chains backfilled to Tools with Conversation mode on. Resource IDs preserved.
  • Conditional Node (no-code) — Field / Operator / Value model with all common operators and AND/OR grouping.
  • Node-level execution conditions — per-step no-code condition with single AND/OR rule mode.
  • Step-level error handling — retry toggle (max + delay), plus four error handling modes: Stop Workflow, Continue with Error, Fallback Value, Fail Branch.
  • Sync mode for Tools — Tool-level toggle for immediate (non-queued) execution; required for Voice-Agent tools.
  • Jinja templating + type casting — {{ }} expressions support conditionals, loops, filters; field-level type casting reduces type-mismatch failures.

Evaluation​

  • New Evaluator setup — LLM-as-a-Judge with Jinja prompt + Python Executor with __main__ signature. No more "evaluator resource" creation or variable mapping.
  • Evaluator templates — user + system templates at the project level. Save evaluators as reusable templates.
  • AI Autofill — generate evaluator prompt or code from a name + description.
  • Multi-version comparison — compare up to 6 versions side-by-side; comparison table with best-value highlighting; "Run on latest version" workflow.
  • Aggregate results + Excel-like table — sortable, filterable, column grouping; color-coded labels; per-evaluator aggregation cards.
  • Scheduled evaluations (cron) — time-based or full cron expressions with timezone + preview; sampling percentage on future traces.
  • Async results export — CSV via email, configurable row range.

Observability​

  • Tracing renamed to Observability — section name change reflected in product and docs.
  • Live tracing — in-progress run view with tree view of completed / failed / in-progress steps, source column (UI/API/Embed), status filters, CSV download.
  • Run History page — KPI cards + Metric Details with per-metric tabs (Runs, Reliability, Latency, Cost, Tokens, Steps, Context Utilization, Tool Calls, Tool-Call Latency).
  • Audit log Phase 2 — side-by-side JSON diff (Old vs New) with red/yellow/green highlighting; 10-minute grouping for draft edits to reduce noise; CSV download with before/after JSON.
  • Audit log UI for CRUD — account-level change history with Project / Application type / Action / Changed by / Date range filters.

Guardrails​

  • Project-level policies — named, reusable Guardrails configurations at the project level. Save from an agent; apply with one click.
  • Guardrails as Tool steps — each guardrail now available as a standalone step in the Tool builder.
  • New validators — Bias Check, Prompt Injection / Jailbreak Detection, Logic Check, Reading Level, Reading Time.

Studio​

  • Auto-titles — newly created applications and conversations get a meaningful title automatically after the first run / first response. One-shot per app/conversation; never overwrites a user-set title.
  • Application deletion rules — must disable a deployed app before delete; must remove from dependents first. Model and connection usage are flagged.
  • Enriched export with resource names — exported configs embed human-readable resource names alongside IDs, so cross-deployment imports show readable missing-dependency lists.
  • Integration tab renamed to Deployment — applied across Agents, Tools, and AI Chains.

Connections & Integrations​

  • Karix WhatsApp — Agent-to-WhatsApp integration via Karix RCM. Optional inbound voice message transcription. Per-message billing with separate ASR line items.

Templates & Marketplace​

  • Internal Marketplace (M1) — Builders share agents within their tenant; Cloners discover, try, and clone. Sandboxed trial sessions billed to the Cloner. Setup dialog auto-opens for missing credentials. Admin controls for who can share / clone.
  • Builder analytics — total clones, time-series, tenant breakdown (no PII), trial counts, listing impressions. Retained after unpublish.
  • KB stubbing via AI — actual KB data isn't shared; Builders can generate dummy KBs via AI prompts.

SDK & Developer​

  • Python SDK (Type I) — workflow execution (sync / async / scheduled / bulk), agent chat (sync + streaming), execution traces, billing data. Authentication abstraction; typed exceptions; configurable timeouts and polling intervals.

Knowledge & Data​

  • Multi-column semantic search — semantic search over multiple structured columns with full support in Tools, AI Chains, and Agents.
  • EdgeQuake RAG framework — internal RAG framework improvements.
  • OCR Parser fix — OCR parser tool fixes for unstructured file processing.

Administration​

  • User active / inactive instead of delete — deactivated users don't count against the plan limit; reactivation is supported.
  • Reinvite — admins can resend invitations to invited-but-not-yet-accepted users when the original invite expires.

Provider additions​

  • OpenRouter + Vercel added as LLM providers.
  • Google added as a STT provider for Voice Agents.

Naming and terminology​

  • AI Chain → unified into Tool in Conversation mode.
  • Tracing → Observability.
  • Integration tab → Deployment tab.
  • (Filename fixes) oolama → ollama, reseach → research, promt-strategy → prompt-strategy.

Documentation​

The entire documentation site has been restructured around 17 top-level sections aligned with user workflows rather than the internal object model. Detailed migration notes for old URLs are tracked in the repo's NOTES_FOR_TECH_TEAM.md.