Follow our journey as we build the future tools for businesses. Here you'll find all updates, new features, and improvements.
The largest internal refactor in Worqs' history — paired with visible improvements across every surface. The Routine engine is rewritten from scratch with three 'kinds' (scheduled, event-driven, reactive), declarative spotter rules, finding merging, a voice runtime, and tight integration with briefings and the Worqs Assistant. A brand-new AI Platform replaces the legacy memory tables with atoms, conversations, entities, and working context — all four agent surfaces now share a unified memory core. The voice agent got a thorough UX redesign with hero cards, multimodal screen content, and a simplified field-collection flow. Compliance cycles, an mTLS proxy for secure outbound integrations, SMS signatures, and lower Telnyx latency round out the release.
Routines now come in three distinct kinds: scheduled (run on calendar), event-driven (fired by state transitions and role-bearers on the entity), and reactive (listen in on active AI conversations). The editor is kind-aware and hides irrelevant sections per kind, with a three-card picker as the new entry point. The list page groups routines by kind and surfaces reactive-coverage hints.
Reactive routines use a declarative rule interpreter that runs over each conversation turn. A generator rewrites natural-language spotters into structured rules with tenant terminology support; rules are stored in the database and regenerated via Inngest. The result is faster and cheaper detection than per-turn LLM calls.
The AI Bar and voice agent now run spotter rules on every user turn and can suggest (or directly launch) a relevant routine when signals are detected. The voice agent additionally has a 'propose-objection' window with a sentiment gate so suggestions land in the right tone. Active routine sessions are exposed through a resume tool in both agents.
A full conversation engine with text runtime, voice runtime, and telephony scaffolding lets routines collect fields, follow up, and close autonomously. The scheduler ticks every five minutes with a tolerance window, Realtime broadcast keeps the list view live, and an admin tab shows sessions and analytics. Briefings now persist routine findings as memory atoms.
Multiple routines sharing the same fire pattern and assignment are automatically merged into a single group in the list view — no more double-pinging the same recipient. Opportunistic spotters can propose a routine launch from an ongoing conversation, and users can start a routine manually at any time.
The old memory and session tables have been replaced by a new AI Platform built on three primitives: conversations (with rolling summary), memory (atoms in six categories, four-level scope: platform → tenant → user → entity), and working context (token-budget-aware prompt assembly). All four agent surfaces — AI Bar, Architect, Voice, and Phone — read and write through the same core. Migration runs as shadow-write for a safe rollout.
Agents can now query memory three ways — by topic, by entity, or on the timeline — and results are fused via Reciprocal Rank Fusion. Atoms can also decay or be archived automatically via a nightly Inngest job, and sensitivity classification is baked into the extraction prompt so private data is handled correctly from the start.
The platform aggregates patterns across multiple conversations — when the same observation recurs across different users or sessions, a rule is proposed that administrators can review and confirm in a new pattern-review UI. Includes recall rate limits, a platform metrics page, and quota enforcement.
The voice page has been rebuilt from scratch: new VoiceInterface shell, seven types of hero cards (list, document excerpt, diagram, image, markdown, form, confirmation), a card carousel instead of overwriting, a new mute pill with tap and 500 ms long-press, a voice menu with voice picker, ambient gradient background, and a status bar with the Worqs logo. Render crashes and hydration mismatches are fixed via layout:false.
A single field-collection flow now handles both issues and business objects. Values for reference fields (user, department, location, BO) are resolved against existing entities, and a dropdown is shown when the match is ambiguous — tap-to-select works for both lists and option fields. Sticky-hero ensures the card doesn't flip back to the welcome screen while the user is thinking.
When a user transitions from voice to text, the handoff carries a one-line summary so the chat agent can pick up exactly where the voice session left off. A per-(tenant, entity) cache for context-delta cuts navigation latency by 50-150 tokens.
Reviews are now organized into version-tracked cycles: create via wizard, collect evidence, submit for approval as a draft, and publish atomically only once approval lands. Cycle tab with a framework activity timeline, skip log for proposed edits, an auditor view, and evidence package export. A 'force cycle' gate prevents publishing outside an active cycle where required.
The compliance consultant has been rewritten from scratch with multi-step planning, single-click framework cards, and updated guidance for the adoption flow. Multi-framework adoption lets several frameworks share the same evidence links simultaneously. A post-adoption prompt reminds the user to start a review cycle.
Outbound integration calls that require mTLS (Nextus, banks, certain government APIs) are now routed through a Cloudflare Worker that handles client certificates for you. Includes a certificate provisioning service, HMAC signing, automatic expiry monitoring, a runbook, and a new mTLS configuration surface in the integrations UI.
Tenant administrators can configure an SMS signature that is automatically appended to outbound messages — including an opt-out hint for GDPR compliance where required.
The first step in a system for tenant-authored detail, list, and relation views. Storage tables, schema versioning, and support for EnterprisePicker, Icon, and Image sources are in place. The live author UI and runtime compiler land in the next release.
The Gemini Live system prompt in the call engine has been densified by 53%, VAD thresholds are trimmed, and post-tool mute is shorter for faster turns. The transfer lifecycle in phone-worker now handles the bridge, initiation phase, and deferred WS-close correctly — outbound transfer legs are no longer hung up prematurely and the Gemini inference meter is stopped when the bridge is up.
Conversation messages are now written via an atomic RPC that removes a TOCTOU race on concurrent writes. A nightly Inngest job provisions next month's partition for the tool-call log so writes never stall on a missing partition.
Closed gaps where Telnyx webhooks could be silently dropped — all events are now logged to `telnyx_webhook_log` and missing events can be traced. The phone number list now shows numbers even when the tenant doesn't have an active subscription.
ClientContent 1007 close on mid-session updates is fixed, duplicate speech from parallel tool calls is deduped, screenContent from multiple tool families (document, business_object, utilization_report) is normalized at the API boundary, and flow tools now get an explicit speakable response so the agent doesn't go silent mid-flow.
The issues page now filters against the correct table — `file_attachment` instead of the non-existent `issue_attachment`. Entity references in memory atoms could go stale on rename, which has also been fixed.
One of the platform's biggest releases: a complete billing engine with EU VAT, metered telephony, and dunning; a hybrid AI router that learns from your queries and now handles 93.7% of Worqs Assistant traffic; the voice agent migrated to Gemini Live 2.5 native audio; the Architect agent can now read the web, draw workflow diagrams, and propose design diffs directly in chat. Plus: bulk import of business objects from CSV/Excel/JSON, the first iteration of AI-driven document assessment, and a substantially improved telephony and messaging stack.
The AI Agent is here: automate entire workflow steps with tool-use, learn from your decisions, and a rule engine that makes the agent smarter with every run. Full visibility through a three-level automation dashboard. Group business objects into categories, render system diagrams directly in your documentation, and several important fixes in the approval flow.
Upload a document, snap a photo, or record a voice memo — AI fills in the form for you. Custom views with resource planning and timesheets. Enterprise scoping isolates data per subsidiary. Gemini Live 3.1 delivers faster voice interaction.
Routines make your AI assistant proactive — it monitors your business objects and issues and reports what's relevant in the morning briefing. Built-in telephony with AI call scripts and automatic transcription. The Architect builds entire systems for you in a multi-agent pipeline.
Major release with a complete compliance framework for ISO and EU regulations, an entirely new AI chat experience with rich cards and tools, and automatic generation rules for calculated fields.
Major release introducing 3D BIM model viewing, customer-facing portals, and an AI voice agent for hands-free interaction.
Major release introducing native approval workflows and a comprehensive system admin dashboard.
Introducing floor plan visualization and 14 new ways to view your business object relationships.
The first public release of Worqs - The Operating System for Your Business.
Subscribe to our changelog and get notified about new features and improvements directly in your inbox.