One Canvas, Way Less Clicking: Inside AI Agent Studio's Redesign

[laszloballa]

If you've built anything in the old AI Agent Studio, you know the drill. Build the agent in one tab. Flip over to check the workflow. Flip again to hunt down the tool you swore you already added. That kind of friction is exactly the feedback ServiceNow says it heard from customers, and it's why AI Agent Studio just got rebuilt from the ground up for Brazil.

@Victor Chen, Senior Principal Outbound Product Manager, walked through the redesign on this week's Developer Passport, then spent the rest of the session building a live agent on an actual instance instead of clicking through slides. No mockups, no rehearsed screenshots, just the real thing.

 

What Actually Changed

The headline is one canvas: agents, tools, and workflows in connected views instead of separate windows you bounce between. If you've used Virtual Agent Designer or any other node-based drag-and-drop tool, it'll feel familiar fast. Under the hood it's run on Lux, our new UI framework.

A couple of other things worth knowing:

• Better integration with AI Agent Advisor (the tool that surfaces automation opportunities from your data and lets you test agents against real records before deploying), mainly around testing and evals.
• Versioning finally covers the whole agent or workflow, not just the instructions text like before. Tools, triggers, everything now versions together.
• Update sets generate in one click from the agent or workflow view. Tools aren't included in that generated set though, so plan to add those separately if needed.

One naming note (unrelated to this redesign): what you knew as "Now Assist" is being renamed "Otto" platform-wide, announced back at Knowledge26. It's not an AI Agent Studio-specific change, but you'll see it bleed into the UI here too.

Availability and Requirements

Entitlement hasn't changed: Pro Plus or an AI SKU. What you need is the latest instance version, Australia Patch 6 or Zurich Patch 12 (both shipped September 10), plus the latest AI Agent Studio store app. Brazil gets it with Patch 0 on September 24, the same day it lands for personal developer instances. The new experience is on by default, just navigate to AI Agent Studio.

Two quick notes: Legacy Studio isn't going anywhere yet, it's one click away in Settings (now bottom left instead of top right). And if you run EmployeeWorks alongside Agent Studio, there are some known compatibility issues being worked on. Log a support case if you hit something weird, chances are it's already tracked.

 

Quick Refresher

Nothing new conceptually, but worth restating since the terms carry the rest of this post: an agentic workflow solves a business problem end to end, and it's made up of one or more AI agents, the more granular units that actually carry tools. Victor's running example was a workflow to triage and categorize ITSM incidents, built from three agents: one linking major incidents, one classifying services, one categorizing the incident itself.

One habit worth adopting: write the description field like you're helping an LLM find it later - because you do. When someone types something loose like "triage this" instead of the exact workflow name, the description, not the title, is what determines whether the right workflow surfaces. Pack in genuinely relevant detail, within reason, and LLM discovery gets noticeably better.

Triggers are optional, and Victor's advice is to use them sparingly: natural language through the panel already works without one. You can also fire a workflow on a schedule or table event, but wiring one to fire on every incident creation is a fast way to turn a helpful automation into a very talkative background job nobody asked for.

 

The Live Build: MCP as a Tool

Victor connected an agent to a free, no-auth MCP server (Microsoft's own docs server) to show how little setup is involved: register the server under Settings, and its tools show up automatically to attach. From there it's a normal agent build, name, description, instructions, tools, access rules, with build and test now sharing one tab instead of two.

One tip worth stealing on its own: name your tools explicitly in agent instructions, and name your agents explicitly in workflow instructions. Saves you and the agent from guessing later.

The full setup, tool config, and test run are worth watching directly in the video above rather than reading a play-by-play here.

 

What's Coming, October and Beyond

• Natural language agent building, a real step up from what exists in legacy Studio today.
• One agent, both modalities. Today you build a chat agent or a voice agent separately. The roadmap moves toward a single build that works across both, instead of maintaining two versions of the same agent.
• Continued improvements across agentic evals, AI Specialists, and AI Agent Advisor itself.

 

Which Model Powers This

Worth knowing: agents can run on Claude, Gemini, Azure OpenAI, or ServiceNow's own LLM (built on an open-weights model), all hosted and billed through ServiceNow. If your org already licenses one of these directly, you can bring your own key, just know you'll then be billed twice, once by the provider, once through your Otto consumption. Bringing an arbitrary open-weights model like DeepSeek or Grok isn't supported yet.

 

Watch the Full Episode

The full session, including the live build and the Q&A, is embedded above. If you want the Developer Passport badge, the window to comment and qualify closes September 25th, so head over to this blog. 

Over to you: what's the first external MCP server you're planning to wire up to an agent, and why that one?

This episode covered forward-looking product plans. Development timelines and feature availability are subject to change, so base any decisions on what's actually shipped rather than what's on the roadmap.