Releases ship weekly. Material changes also go to customers by email.
Agents can now route synthesis and transcription through ElevenLabs, OpenAI, or Azure Speech, configurable per agent. The default fallback chain prefers the lowest-latency healthy provider in your region.
The voice console replaces the separate start, mute, and end buttons with a single state-aware control. Press once to talk, press again to end — barge-in stays on by default.
Intent routing now compares queries against intent corpora using a normalized token-bag similarity score with stemmed n-grams. We measured a 19% improvement in routing accuracy on the noisy-input benchmark.
If the cloud TTS provider is slow or unreachable, the agent now falls back to the local Web Speech voice without a visible error — the reply continues mid-sentence instead of stalling.
Holds the microphone open between turns, detects end-of-utterance with a tunable silence threshold, and resumes listening as soon as the agent stops speaking. Disabled by default; enable per agent in voice settings.
Point an agent at a sitemap or a list of URLs and we'll crawl, chunk, and embed the content nightly. Retrieval is scoped per agent so workspaces sharing a tenant can keep their corpora separate.
When an agent runs a web query, the top results are now passed to the model as cited context with source URLs included in the response. Hallucinations on long-tail factual questions dropped meaningfully in internal evals.
When the intent router is below confidence threshold, agents now draft a reply from broader context rather than returning a canned "I'm not sure." Drafts are flagged in transcripts so they can be reviewed and added back to the corpus.
A single workspace can now host more than one production agent under a single billing account. Permissions, knowledge bases, and voice configs are scoped per agent; cross-agent handoff is opt-in.
Connect a Notion, Confluence, or Google Drive source and we'll surface candidate documents grouped by likely topic. Approve, edit, or reject the suggestion before it joins the live corpus.
Demo agents now ship with English, Spanish, and French presets — including voice, transcripts, and sample knowledge — so you can show the same flow to a regional buyer without rebuilding it.
Thumbs-up and thumbs-down signals now feed back into routing and corpus ranking on a per-agent basis instead of a tenant-wide pool. Noisy reviewers on one agent no longer pollute the others.
Each turn is scored on a -1 to +1 axis and surfaced in the live transcript. Thresholded events can trigger handoff to a human, escalation in the CRM, or a callback flag.
We closed our second-year Type II audit with no exceptions. The latest report is available under NDA from your CSM or by request at security@wonderfulagent.example.
Workspaces can now be pinned to our europe-west4 region. Customer Data — including embeddings, transcripts, and exports — stays inside the EU; inference traffic is routed to in-region model deployments where available.
Stream the workspace audit log to S3, GCS, or any S3-compatible bucket on a 15-minute cadence. JSONL format, including agent actions, admin changes, and access events.
Provision and deprovision users and groups from Okta, Entra, and Google Workspace via SCIM 2.0. Role mappings are configurable per identity provider.
The voice selector dropdown was returning an empty list on Firefox because the speechSynthesis voices event fired before the panel mounted. We now re-query on mount and on focus.
Connect your tools, brief your first agent, and ship work today. Free for 14 days, no card needed.
