Skip to main content
​
SuperFlowAgentsModels and costKnowledge Base
Studio 2.0

​
Studio 2.0

Studio 2.0 lets you build a SuperFlow by describing it in plain language, adds new nodes and a memory module to SuperFlow, lets an agent carry multiple global contexts, and adds side-by-side model comparison to the Playground. The release contains 13 new updates and 3 previously released updates across four areas.Release date: October 26, 2026.The Build with AI panel in SuperFlow. Under the heading "What should this SuperFlow do?", it explains that the builder draws the flow on the canvas and that + attaches an agent, knowledge base, guardrail policy, SuperFlow, or tool. Three example prompts sit under Try asking, above a message box that reads "Describe the workflow you want."

​
New in SuperFlow

Build a SuperFlow with AI. You can now create a SuperFlow by describing what it should do in plain language. Open Build with AI, describe the goal in a sentence or a paragraph, and SuperFlow adds the nodes to the canvas along with a plain-language summary of what the flow does. Select Run sample input to test the flow end to end before you deploy it. To change the flow, describe the change in the same chat. You can undo any change with Ctrl+Z. See SuperFlow.Redesigned SuperFlow canvas. SuperFlow has updated node visuals and a redesigned canvas. The node library can now be searched by name or browsed by category, such as AI, Safety, and Control Flow. See SuperFlow.Knowledge Base and Loop Until nodes. SuperFlow adds two nodes for more complex orchestration on a single canvas. The Knowledge Base node retrieves answers from a Knowledge Base at any step in a flow. The Loop Until node sends execution back to an earlier step and repeats the steps in between until a condition you define is met, or until an iteration or token budget runs out (10 iterations by default). See SuperFlow nodes.Decide and Rank nodes. The Decide node asks a model for a choice, a score, or a probability, using the Jev decision model from Typesafe AI. The Rank node reorders a list of items by relevance to a goal, using the same model. See Decide and Rank.Memory module for SuperFlow. A SuperFlow can now remember what it has already processed. Set Memory on an AI Agent or LLM node to Per execution to keep memory within a single run, Across executions to keep it from one run to the next, or Custom key to group memory by a value you supply, such as a customer ID. The flow uses what it has already seen to decide its next step. This setting does not turn off memory already configured on the agent. See AI Agent.Check, save versions, and deploy. SuperFlow now keeps your edits in a draft and runs production traffic from a deployed version. Select Check to find problems before you run, save the draft as a numbered version, and deploy it when it is ready. Webhooks, schedules, and API calls run the live version, and you can roll back by deploying an older one. See Check, save, and deploy.Apps in SuperFlow. Each app is now its own node under Apps in the node library, including Gmail, Slack, Google Docs, SharePoint, and Microsoft Outlook. Connect your account on the node, then pick the action to run. See App tools.Create agents inside SuperFlow. You can now create an agent, configure it, and connect it to a flow without leaving the SuperFlow canvas. See SuperFlow.

​
New for agents

Multiple global contexts per agent. An agent can now use up to 10 global contexts, applied in the order you list them. For example, you can keep brand voice, product facts, and policies as separate global contexts and attach each one to the agents that need it. See Global Context.Create global contexts and skills from the agent builder. You can now create global contexts and skills on a single page inside the agent builder, without switching to another screen during a build. See Studio Builder and Skills.

​
New in models and cost

Compare models side by side in the Playground. The Playground can now send one message to two or three models at once and show their responses next to each other. Select Compare models in the chat, pick the models, and send your message. Each model answers with the same instructions, knowledge, and context. Each response shows its token count, total response time, and time to first token (TTFT). The Playground labels the response that used the fewest tokens and the response that arrived fastest. To switch an agent to one of the compared models, select Use this model under that response. See Playground.The Agent Studio Playground comparing three models (gpt-5.6-luna, claude-sonnet-5, and grok-4.6) on the same prompt in side-by-side columns. Under each response, a row shows the token count, total time, and time to first token. The first response is labeled Fewest tokens and marked In use, the second is labeled Fastest, and the other two have a Use this model button.

​
New for Knowledge Base and teams

Folder-level SharePoint sync. You can now choose which SharePoint folders a Knowledge Base syncs, instead of syncing the full connected location. See Knowledge Base Studio.Share skills with team members. You can now share a skill with one or more members of your team, with view or edit permission, so they can use it in their own agents. See Skills.

​
Previously released

The following updates shipped before Studio 2.0 and are available in Agent Studio today.Jev decision model. Jev is a classification-only model from Typesafe AI. In SuperFlow, the Decide node uses Jev where a step needs to classify input rather than generate text. You can also choose Jev as the request classifier for auto-routing. See Models.Auto-routing by complexity. When you turn on auto-routing for an agent, Agent Studio classifies each request by complexity and serves it from the model you assigned to that tier, instead of the model selected on the agent. Each tier shows its model’s cost and speed, so you can see the trade-off before you assign a model. First released on August 21, 2026. See Models.Rate limits at four levels. You can set rate limits for an organization, a sub-organization, an agent, or an individual user. An organization limit sets one ceiling for all usage, a sub-organization limit sets a budget per team or client, and agent and user limits apply to a single agent or person. First released on August 21, 2026. See Rate Limits.

​
Updates before Studio 2.0

The updates below shipped before Studio 2.0. They are listed by date, newest first.
​
  • Auto-route by complexity: Classify each request by complexity and serve it from a per-tier model (Simple, Medium, Complex, Reasoning) instead of the model selected on the agent. Cheap, fast models handle simple requests automatically while harder ones still reach a stronger model.
  • Rate limits for users, agents, and sub-organizations: Cap credits and/or tokens over a rolling window (hourly, daily) for individual users, agents, or sub-organizations, on top of space-level credit limits.
  • File upload trigger for SuperFlow: A new SuperFlow trigger mode for document-processing pipelines. The workflow receives file_url, file_name, file_type, and file_size as input fields. Pair it with a Parse node to extract the file’s contents.
​
  • Admin Dashboard Insights: Enterprise plan admins can now view all user insights in a single dashboard, making it easier to spot trends and make decisions across the organization.
  • Custom Guardrails: Connect a guardrail to your own HTTP endpoint to enforce organization-specific validation, compliance, or moderation rules.
  • Improved Git integration: Connecting an agent to Git now supports a lot more functionality.
  • One-Shot RAG: Retrieves relevant context and generates a response in a single LLM call, ideal for straightforward question-answering where one retrieval pass is enough.
  • A2A in the registry: Agent-to-Agent (A2A) agents can now be added directly to the registry, in addition to being used in orchestration.
  • New memory providers: Added GBrain, Azure AI Foundry, Vertex AI Memory Bank, and Vertex AI Sessions.
  • Voice: Bring Your Own Model: Voice agents now support bringing your own model.
  • Agent tagging and deactivation: Add tags and metadata to agents for easier filtering, and deactivate any agent when it’s no longer needed.
  • Simpler tool selection: When adding tools, all actions are selected by default instead of requiring you to pick each one, making agent building faster via both Architect and the conversation builder.
  • Usability improvements:
    • In-product feedback collection.
    • Playground fixes.
    • A utility agent that helps you pick the right model for your agent.
    • A utility agent that analyzes your traces.
​
  • Studio redesign: A cleaner design, improved navigation, and a more intuitive experience across Studio. The sidebar is better organized, and core flows like Agent Builder, Knowledge Base, Playground, and Deployment are now faster to use.
  • Agent Improvement Engine: Keep production agents reliable by monitoring performance, catching issues early, receiving proactive alerts, and setting rate limits to prevent unexpected costs.
  • Usage reports: Download detailed usage reports for visibility into agent performance across your organization, including usage trends, key metrics, and per-user activity.
  • Guardrails node in SuperFlow: Add guardrails as a node directly in your SuperFlow workflows.
  • ArangoDB connector: Added as a new Data Connector.
  • Storage security: Improved security for all stored files, including file uploads, Knowledge Base, and Knowledge Graph.
​
  • Native app integrations: Surface any Lyzr agent in Microsoft Teams, Slack, or Telegram and talk to it directly from the app your team already uses.
  • Vertex AI models: Now available on the Models page, where you can add your own credentials.
  • Fallback models: Configure up to 3 fallback models so your agents keep running automatically if an LLM provider goes down.
  • New guardrails: Valid OpenAPI Schema and Valid SQL, to validate SQL queries and OpenAPI schemas directly within your agents.
  • AI Tool Suggestor: Describe what you need in plain English while adding tools, and Lyzr identifies the right tool and capabilities for you.
  • PII Guardrail improvements: Better detection and masking quality for personally identifiable information.
  • SharePoint permission sync: Knowledge Base access now respects SharePoint folder and file permissions.
​
  • Skills: Modular building blocks that define what an agent can do, including code, tool integrations, or multi-step prompts. Skills can interact with resources like filesystems or external APIs to enable structured, repeatable, complex workflows.
  • Near real-time Knowledge Base: Added Microsoft SharePoint support, with Google Drive coming soon. Webhook-powered updates pull the latest information from SharePoint files, and tracked websites update automatically when they change.
  • Sub-organization credit limits: Set credit limits for sub-organizations and get an email notification whenever one is about to cross its limit.
  • New memory providers: Added Mem0 and Supermemory.
  • LLM-as-Judge: Validate agent responses using an independent model. Choose the model, define custom evaluation instructions, and let an external agent review outputs.
  • New database connectors: Added Oracle and Databricks.
  • Grok models: Now available as a model provider.
  • CSV as direct file input: Upload CSV files directly as input, with no extra conversion step.
  • MCP resources and prompts: MCP integrations can now bring in resources and prompts, not only tools.
  • Gibberish Text guardrail: New guardrail to detect gibberish text.
  • Improved resource sharing: Sharing an agent now shares its connected Knowledge Base too, and sharing a manager agent shares all of its sub-agents.