changelog: September and October product announcements

docs.json

@@ -560,6 +560,7 @@
                 "group": "Governance & Controls",
                 "pages": [
                   "enterprise/integration-controls",
+                  "enterprise/model-controls",
                   "enterprise/default-org",
                   "enterprise/manage-projects-and-users-api",
                   "enterprise/asset-controls",

site/content/docs/changelog/index.mdx

@@ -4,6 +4,118 @@
 mode: "wide"
 ---
 
+<Update label="October 30, 2026" description="Model controls for organization admins">
+
+## Model controls for organization admins
+
+![Models tab of the Integrations & models settings page, with new models set to Block all and the Anthropic and Google models expanded, where Claude Haiku 4.5 is switched off and its Reading web pages for research feature is marked as stopped](/images/changelog/changelog-model-controls.png)
+
+Organization admins on Enterprise plans can now choose which models their organization can use, and see what a change affects before it applies.
+
+Switch off a single model or a whole provider, and choose whether new models start on or stay off until an admin allows them. Each model lists the platform features it powers, such as memory and web reading, and before you confirm you see which features stop or work with limits. The confirmation also warns that your own Agents and LLM tool steps on that model stop until you switch them to another model.
+
+Open organization settings, go to Integrations & models, and select the Models tab. See [Model controls](/enterprise/model-controls) for the full guide.
+
+</Update>
+
+<Update label="October 20, 2026" description="Error, approval, and escalation alarms added to Tasks page">
+
+## Error, approval, and escalation alarms added to Tasks page
+
+![Tasks page on the Approvals view grouped by Agent, with checkbox, Alarm, Run by, Tasks, and Task approval rate columns, where Invoice approver has its alarm turned on at an 18.0% task approval rate](/images/changelog/changelog-task-alarms.png)
+
+You can now set alarms on an Agent or Workforce that notify you when too many of its tasks fail, sit waiting on an approval, or get handed off to a person.
+
+Each alarm fires when the share of tasks crosses your threshold in a time window, for example more than 8% within an hour, once at least 10 tasks have run. Alerts go to email, Slack, or Teams, with a follow-up when things recover. Turn on Auto-diagnose with Invent and Invent proposes a fix for you to approve. Task alarms are available on paid plans.
+
+To set one up, open the Tasks page, switch to the Errors, Approvals, or Escalated view, and group by Agent or Workforce. Click the bell in the Alarm column for a single row, or select several rows and click Enable alarms. See [Tasks page](/build/agents/give-your-agent-tasks/tasks-page) for what each view shows.
+
+### GPT-6 Luna Decisions now available
+
+OpenAI's GPT-6 Luna Decisions joins Jev 1.13 in Workforce condition nodes and the Decide tool step. It runs without your own API key in the US and EU. See [Add conditions](/build/workforces/build-an-ai-workforce/add-conditions) for how condition nodes route a Workforce.
+
+### Claude Sonnet 5.5 and Claude Haiku 5.5 now available
+
+Sonnet 5.5 balances speed and intelligence for coding, Agents, and knowledge work. Haiku 5.5 is the fastest, lowest-cost Claude model, suited to high-volume classification, extraction, and routing. Select either from the model dropdown in any Agent, Tool, or Workforce. See [language model settings](/build/agents/build-your-agent/agent-settings/language-model) for choosing an Agent's model.
+
+</Update>
+
+<Update label="October 2, 2026" description="Decision models added to Workforce conditions and Tools">
+
+## Decision models added to Workforce conditions and Tools
+
+![Condition panel in the Workforce builder with Evaluate prompts with set to Decision model and Jev 1.13 selected, running on the TypeSafe key, above a prompt-based condition that routes to a refunds Agent](/images/changelog/changelog-decision-models.png)
+
+Jev 1.13, a decision model from TypeSafe, is now available in Workforce condition nodes and in a new Decide tool step.
+
+A decision model answers multiple choice, score, and yes/no questions with a probability for every option, in under a second. It doesn't write text, so it suits routing, triage, and qualification.
+
+Add your TypeSafe API key on the Integrations page. Then, in a Workforce condition node, set Evaluate prompts with to Decision model and pick Jev 1.13, or add the Decide step to a Tool. See [Add conditions](/build/workforces/build-an-ai-workforce/add-conditions) for how condition nodes route a Workforce.
+
+### MCP server access in Invent
+
+Invent can now call tools on your project's MCP servers, such as GitHub, Linear, or Notion, and you can approve a tool for a single chat for 24 hours. Turn it on in Invent settings under MCP servers.
+
+### Dark mode now available
+
+Relevance AI now has a dark theme across the app, including Invent. Click your profile at the bottom of the left navigation menu and choose Appearance, then pick Match system, Light, or Dark. The choice is stored in your browser, so it applies per device.
+
+</Update>
+
+<Update label="September 25, 2026" description="Trigger status page and failure alarms">
+
+## Trigger status page and failure alarms
+
+![Trigger status page listing triggers with alarm bells, the Agent or Workforce each belongs to, statuses of Failed, Paused, Cancelled, and Active, and last successful and last failed run times](/images/changelog/changelog-trigger-status.png)
+
+The new Trigger status page lists every trigger in a project with its current status and latest runs, and can notify you, or Invent, when one fails.
+
+Each row shows the trigger, its Agent or Workforce, its status, and its last successful and failed runs. A failure alarm alerts email, Slack, or Teams when a trigger fails and again when it recovers. With Auto-diagnose with Invent on, Invent investigates each failure and proposes a fix for you to approve.
+
+Click Triggers in the left navigation menu to open the page. To set an alarm, click the bell on a trigger's row, optionally turn on Auto-diagnose with Invent, and click Enable alarm. Auto-diagnosis needs Invent turned on for the project. See [Triggers](/build/agents/build-your-agent/triggers) for how triggers start tasks.
+
+### Diagrams in Invent
+
+Ask Invent to visualize something, such as the branching logic in an Agent's system prompt, and it draws a Mermaid diagram in the chat, with Diagram and Code tabs, a Copy button, and an expanded view with zoom.
+
+### CSV bulk runs up to 10,000 rows
+
+Bulk run CSV uploads now handle up to 10,000 rows without timing out. Runs show Preparing while rows load and Failed if no tasks ran, instead of Complete with 0 of 0 tasks. See [Bulk scheduling](/build/agents/give-your-agent-tasks/bulk-schedule) for how to set up a bulk schedule from a spreadsheet.
+
+</Update>
+
+<Update label="September 17, 2026" description="Scheduled routines in Invent">
+
+## Scheduled routines in Invent
+
+![Routines panel in Invent with Yours and Templates tabs, a Schedule with Invent button, and six routines such as Reflect, Sharpen instructions, Cost optimize, and Check the outbound queue with their daily and weekly schedules](/images/changelog/changelog-invent-routines.png)
+
+Invent can now run work on a schedule, such as reviewing yesterday's Agent runs or compiling a weekly report from connected tools.
+
+Ask Invent in chat to do something on a recurring basis and it creates a routine that runs daily, weekly, monthly, or annually. The Routines panel lists each routine with its schedule, and the Templates tab offers ready-made routines to start from.
+
+Click Routines in the Invent sidebar to see and edit your routines, or click Schedule with Invent to set up a new one.
+
+### Status feed added to Home page
+
+The Home page now shows a 14-day task health bar and performance for up to five Agents or Workforces. Click the gear on the Status feed to choose which ones appear. It's visible to Admins and Editors.
+
+</Update>
+
+<Update label="September 11, 2026" description="Subagents in Invent">
+
+## Subagents in Invent
+
+![Invent chat splitting a request across four subagents, each with its own mascot: two working with open eyes, one done with happy eyes after 9 steps, and one queued and asleep](/images/changelog/changelog-invent-subagents.png)
+
+Invent can now hand independent parts of a task to subagents that run in parallel.
+
+When a request has independent pieces, such as researching leads while building an enrichment Tool, Invent starts a subagent for each and tracks them as cards in the chat. Subagents can't publish, delete, or restore anything, so those actions still go through your usual approvals.
+
+Subagents start automatically. To see what each one cost, click the credits button in the composer; the breakdown lists each subagent as its own row. See [Build with Invent](/build/invent/invent) for how Invent's credit counter works.
+
+</Update>
+
 <Update label="September 3, 2026" description="Meet the new Invent">
 
 ## Meet the new Invent
@@ -12,7 +124,7 @@
 
 The newest version of Invent is now available.
 
 Invent can create and edit Agents, Tools, Workforces, Knowledge tables, Triggers, and Evals, run tasks, pull your analytics, and much, much more. It has context over everything in your project and acts as a specialist who can help you create almost anything inside of Relevance AI.
 
 Open Invent from the badge tucked into the right edge of the screen, from the button in the bottom-right corner of the Home page, or from Invent in the left navigation menu. Invent is enabled per project: an organization admin opens Settings → Projects, selects the project, clicks Manage settings, and turns on the Invent toggle. See [Build with Invent](/build/invent/invent) for more information on using and configuring Invent.
 
@@ -22,27 +134,27 @@
 
 ## Concurrency visibility for project and organization admins
 
 ![Concurrency panel showing a seven-day usage chart with queued runs stacked in red above the limit line, 20 of 20 slots in use with 6 tasks waiting, and a live capacity by project bar with other projects anonymized](/images/changelog/changelog-concurrency-visibility.png)
 
 Concurrency monitoring on the Analytics page now covers queued runs, live slot counts, and a per-project capacity breakdown.
 
 Queued runs now stack in red above the limit line on the usage chart, live counters show slots in use and tasks waiting, and a per-project bar attributes running slots to the projects holding them. Between them you can see whether you're at the limit, how much is queued behind it, and which project is consuming the capacity. Project admins see the same charts with other projects' names and task details anonymized.
 
 Project admins find the chart on the Analytics page; organization admins can also view it on the Plan & Billing page.
 
 </Update>
 
 <Update label="June 29, 2026" description="Eval cost breakdowns and Workforce evaluations">
 
 ## Eval cost breakdowns and Workforce evaluations
 
 ![Eval run cost breakdown panel showing total credits and actions, with per-component rows for Scenario Runner, Agent Execution, and Checks broken down by AI Judge model and Tool Usage](/images/changelog/changelog-eval-cost-breakdown.png)
 
 Eval runs now show exactly where their cost goes, itemized per check, and cover full Workforce runs including sub-agent and tool calls.
 
 Each completed Eval run renders a cost breakdown panel that itemizes credit and action consumption per component. Checks account for 1 action per run across every check type, including LLM as Judge and Tool Usage, so the panel's totals reconcile to the exact source of every credit and action. Workforce runs that fan out to sub-agents or call tools record and evaluate each of those interactions, the same as individual Agent runs.
 
 To see the breakdown, open any completed Eval run; the panel appears alongside the results. See the [Evals documentation](/build/agents/evals/introduction) for full details.
 
 </Update>
 
@@ -80,7 +192,7 @@
 
 ![Microsoft Teams trigger settings showing the new Outreach replies only toggle enabled](/images/changelog/changelog-teams-outreach.png)
 
 Microsoft Teams triggers now support an **Outreach replies only** mode that limits responses to direct replies on the Agent's own outbound DMs.
 
 The trigger distinguishes between inbound replies (replies to messages your Agent sent) and channel mentions. In outreach-only mode, only inbound replies route to the Agent — channel chatter doesn't trigger a run.
 
@@ -134,7 +246,7 @@
 
 You can now manage public sharing settings for all agents, workforces, and tools from one central location!
 
 Project-Level Public Agent Controls give admins the power to disable public cloning and embeddability across their entire workspace with a single setting. When you disable these features at the project level, they’re automatically turned off for all agents, workforces, and tools—and individual users can’t re-enable them.
 
 ➡️ Centralized control – Manage security settings for all assets from one place
 
@@ -182,7 +294,7 @@
 
 ![Agent Performance Observability](/images/changelog/changelog-observability.webp)
 
 You can now track how your agents perform in real-world scenarios with comprehensive production monitoring!  Just like software development has build → test → monitor stages, agent development now has the same complete lifecycle. After building your agents and testing them with Evals, you need visibility into production performance at scale.
 
 ➡️ **Automatic evaluation** – Set up LLM judges to continuously assess your agents on custom criteria like summary quality, hallucination detection, and workflow execution
 
@@ -194,7 +306,7 @@
 
 With Agent Performance Observability, you can catch performance degradations immediately—like when a model update changes behavior or agents start hallucinating—instead of discovering issues weeks later.
 
 🔐 **Permissions required:** Available exclusively for select Enterprise customers. Evals delivers the advanced testing capabilities needed for mission-critical AI deployments. This feature is not available on non-Enterprise plans.
 
 To access Agent Performance Observability, enterprise users can go to their agent’s dashboard and select the “Monitor” tab, then navigate to the “Test Suites” section. If you’re on an Enterprise plan and don’t yet have access, contact your Relevance AI sales representative to register interest.
 
@@ -253,13 +365,13 @@
 
 </Update>
 
 <Update label="March 10, 2026" description="Monitor organization concurrency usage with detailed analytics and timeseries data">
 
 ![Monitor organization concurrency usage with detailed analytics and timeseries data](/images/changelog/changelog-concurrency.webp)
 
 **Org Concurrency Timeseries: Track Your Organization’s Performance Analytics**
 
 You can now monitor your organization’s concurrency usage and limits with detailed timeseries data!
 
 View comprehensive analytics directly from your organization settings page under the Plan & Billing section, giving you complete visibility into how your team uses Relevance AI’s resources.
 
@@ -269,9 +381,9 @@
 
 ➡️ Usage limit monitoring – Keep track of your current usage against your plan limits to avoid unexpected throttling
 
 ➡️ Visual timeseries charts – Easy-to-read graphs show your concurrency patterns, helping you identify optimization opportunities
 
 With Org Concurrency Timeseries, you can make data-driven decisions about your AI operations, ensuring optimal performance and resource allocation across your entire organization.
 
 🔐 Permissions required: Organization Admin access to view billing and usage analytics
 
@@ -507,11 +619,11 @@
 
 ![Claude Opus 4.6 model now available in Relevance AI](/images/changelog/68faec2d3ffbd34e9c9295b9-changelog-claude.webp)
 
 **Claude Opus 4.6: Anthropic's Most Advanced AI Model Now Available**
 
 You can now access Anthropic's flagship Claude Opus 4.6 model directly in the platform, with an optional 1M context window variant for handling massive documents!
 
 Claude Opus 4.6 delivers exceptional performance on complex reasoning, coding, and advanced agentic tasks. This latest version brings Anthropic's most powerful AI capabilities to your workflows, with both standard and extended context options to match your specific needs.
 
 ➡️ Standard 200K token context window – Process lengthy documents and conversations with ease
 
@@ -519,7 +631,7 @@
 
 ➡️ Adaptive thinking support – Tackle complex reasoning and multi-step problem solving
 
 ➡️ Advanced image processing – Work with jpg, jpeg, gif, webp, and png files
 
 ➡️ Exceptional output quality – Benefit from improved reasoning and response generation
 
@@ -563,15 +675,15 @@
 
 </Update>
 
 <Update label="February 9, 2026" description="Test and validate AI agents with Evals: Enterprise-grade agent testing">
 
 ![Test and validate AI agents with Evals](/images/changelog/698a6f9d99180607d34d642c-changelog-evals.webp)
 
 **Evals: Test and Validate Your AI Agents with Confidence**
 
 You can now thoroughly test your AI agents with our powerful Evals feature, ensuring they perform exactly as expected in real-world scenarios!
 
 Evals provides a structured testing framework for your AI agents, allowing you to create test suites with multiple scenarios to validate functionality. Create golden test examples, simulate user interactions, and evaluate agent performance without triggering actual tool executions or external actions.
 
 ➡️ Create organized test suites – Group related tests together for better organization and management
 
@@ -583,9 +695,9 @@
 
 ➡️ Coming soon: Tool simulations – Test workflows with sensitive or expensive tools without triggering actual executions
 
 With Evals, you can confidently deploy agents knowing they've been thoroughly tested against a variety of scenarios, ensuring reliability and consistent performance for your users.
 
 **Available exclusively for select Enterprise customers**, Evals delivers the advanced testing capabilities needed for mission-critical AI deployments. To access Evals, enterprise users can go to their agent's dashboard and select the "Monitor" tab, then navigate to the "Test Suites" section. If you’re on an Enterprise plan and don’t yet have access, contact your Relevance AI sales representative to register interest. This feature is not available on non-Enterprise plans.
 
 Start building more reliable AI agents with comprehensive enterprise-grade testing today!
 
@@ -665,11 +777,11 @@
 
 Workforce Trigger Controls give you granular management over your automation workflows. Quickly pause specific triggers when issues arise, cancel problematic workflows, or pause an entire workforce at once—then resume operations when you're ready.
 
 ➡️ Pause individual triggers – Temporarily stop specific automations without affecting others
 
 ➡️ Cancel active workflows – Immediately terminate problematic processes
 
 ➡️ Resume paused triggers – Restart automations when issues are resolved
 
 ➡️ Pause entire workforces – One-click solution to halt all triggers during maintenance
 
@@ -783,11 +895,11 @@
 
 ![Control your meeting agent in real time with chat commands](/images/changelog/695a95a2a931a05c425b7b72-changelog-meeting-bot-real-time-commands.webp)
 
 **Meeting Agent Realtime Commands: Control Your AI Agent During Live Meetings**
 
 You can now trigger your AI meeting agent with custom commands during live meetings without interrupting the flow of conversation!
 
 Meeting Agent Realtime Commands lets you interact with your meeting bot during meetings by simply typing trigger words in the chat. Your agent springs into action when it detects these keywords, performing tasks like searching information, controlling recordings, or creating meeting artifacts—all while your meeting continues uninterrupted.
 
 ➡️ Use custom trigger words – Define specific phrases that activate your agent during meetings
 
@@ -799,7 +911,7 @@
 
 ➡️ Extend with any Relevance AI capability – Connect to Notion, Slack, or other tools for seamless workflows
 
 With Meeting Agent Realtime Commands, you can enhance your meeting productivity by leveraging AI assistance exactly when you need it, without disrupting the conversation flow.
 
 To set up your own meeting agent with custom commands, create a new agent from scratch, add the "Relevance AI Meeting Bot" under Triggers, add and configure your desired meeting bot tools (such as 'Pause Recording' and 'Resume Recording') and more general tools (such as 'Google Search' and 'Slide Builder') and trigger words.
 

site/content/docs/enterprise/meta.json

@@ -12,6 +12,7 @@
     "user-level-authentication",
     "---Governance & Controls---",
     "integration-controls",
+    "model-controls",
     "default-org",
     "manage-projects-and-users-api",
     "asset-controls",

site/content/docs/enterprise/model-controls.mdx

@@ -0,0 +1,65 @@
+---
+title: "Model controls"
+sidebarTitle: "Model controls"
+description: "Choose which models your organization can use, and see what a change affects before it applies"
+---
+
+Model controls let organization admins decide which models people in the organization can use. You can block a whole provider or a single model, decide whether models added to Relevance AI later start on or off, and see which features a change affects before you confirm it.
+
+<Info>Model controls are available on Enterprise plans. You need to be an organization admin to change them.</Info>
+
+## Open model controls
+
+<Steps>
+  <Step title="Open organization settings">
+    Open your organization settings and select **Integrations & models**.
+  </Step>
+  <Step title="Open the Models tab">
+    Select the **Models** tab. The count beside it shows how many models are on, for example **17 of 19**.
+  </Step>
+</Steps>
+
+## Choose which models are allowed
+
+The dropdown at the top of the Models tab sets every model at once:
+
+- **Allow all** turns every model on, including new ones.
+- **Block all** turns every model off. Agents and Tools using them stop until a model is switched back on, and new models are off by default.
+- **Custom** lets you choose provider by provider and model by model. Your saved choices come back whenever you switch to Custom.
+
+Each provider, such as Anthropic or Google, has its own count and its own **Allow all**, **Block all**, or **Custom** dropdown. Expand a provider to switch individual models on or off with the toggle on each row.
+
+If an organization API key that a model relies on is switched off, the model shows **API key off** and is blocked until the key is turned back on. See [Organization- and project-level API keys](/enterprise/org-project-level-api-keys) for managing those keys.
+
+## Control new models
+
+**Turn on new models automatically** decides what happens when Relevance AI adds a model:
+
+- **Allow all** turns new models on for every provider.
+- **Block all** keeps new models off until an admin turns them on.
+- **Custom** lets each provider decide, using the **Turn on new [provider] models automatically** toggle inside that provider.
+
+Organizations that only allow approved vendors can use **Block all** so models released later never turn on without review.
+
+## See what a change affects
+
+Relevance AI runs some platform features, such as memory, web reading, and error explanations, on models it chooses. Each model lists the features it powers under **Powers**. A red pill marks a feature that your current settings have turned off.
+
+When your settings affect any features, a banner at the top of the tab shows how many, and how many stop or work with limits. Expand it to see each feature and the models it needs.
+
+Before a model turns off, a confirmation lists what it affects:
+
+- **Stops** marks a platform feature that won't run without the model.
+- **Works with limits** marks a feature that keeps running with less capability.
+- **Uses [model] instead** marks a built-in Agent that moves to another allowed model.
+- **Your agents and tools** lists your own Agents and LLM tool steps on that model, which stop until you switch them to another model or turn the model back on.
+
+Turning the model back on restores everything the confirmation listed.
+
+## Frequently asked questions (FAQs)
+
+<AccordionGroup>
+  <Accordion title="What happens to Agents that use a model I block?">Agents and LLM tool steps that use a blocked model stop when they run. Switch them to an allowed model in their settings, or turn the model back on.</Accordion>
+  <Accordion title="Do I lose my choices if I switch to Allow all or Block all?">No. Your per-model choices are saved, and they come back when you switch to Custom.</Accordion>
+  <Accordion title="Can I use model controls without Enterprise?">No. Model controls are available on Enterprise plans only.</Accordion>
+</AccordionGroup>