Brazil reached Early Availability on September 24, 2026, and general availability is expected in November. We have spent the past few weeks reading the release notes end to end, sitting through the Developer Passport sessions, and running the new tooling on a Brazil PDI. This is the short version.
What makes this release different is where ServiceNow is putting the platform's center of gravity: AI that does the work rather than suggests it, governance to keep that AI on a leash, and a developer toolchain that treats AI-generated and hand-written code the same way. Those three threads run through almost everything below.
01
Autonomous Engineer sits on top of Build Agent in ServiceNow Studio and works in two phases. In the plan phase you hand it a requirement, either typed or as an uploaded document. It asks clarifying questions, checks your instance for tables, roles and catalog items it can reuse, and produces a plan of work items written as user stories with acceptance and test criteria. Nothing is built until a person approves that plan.
In the build phase it spins up a background agent per work item, builds them in parallel, generates and runs ATF tests for each one, and attempts to fix its own test failures. Anything it cannot resolve is flagged for review. When the plan is complete, the output is an update set for your UAT or production pipeline. It does not deploy to production on its own.
Build Agent itself also changed. The trial version is now installed and on by default on every instance, it receives automatic upgrades from the Store on Brazil Patch 0, Australia Patch 5 and Zurich Patch 12 or later, and it now supports custom rules (standing guardrails such as naming conventions and scope restrictions) and custom skills you can invoke in a session. Autonomous Engineer inherits both.
Before you turn it on
02
Arc was announced on October 21 and shows up in the Brazil release notes under its working name, ServiceNow Cowork. It is a long-running personal agent that runs on an employee's Mac (Windows is promised within weeks) or inside an NVIDIA OpenShell runtime. You give it a goal rather than a set of steps; it plans, works across ServiceNow, email, Teams and files, checks its own work, and keeps going until it is done, whether that takes minutes or days.
What makes it enterprise-grade rather than a consumer gadget is the control surface. Administrators set policies with network allowlists, file-type rules, connector scopes and tool gates, targeted to users and groups by user criteria. Approvals can be hard (ask every time) or soft (allow once, or always), and an optional auto-approval classifier can clear routine soft-gate calls against the user's stated goal. Arc registers as a managed AI system in AI Control Tower, which gives you fleet visibility, heartbeat monitoring and a kill switch that revokes credentials and terminates running actions.
Commercially, it is included at no extra cost with the Foundation, Advanced and Prime tiers of EmployeeWorks. Model choice spans Anthropic Claude, OpenAI GPT and Google Gemini, routed through the Generative AI Controller, and user-initiated requests consume assists like everything else.
Before you turn it on
03
The inventory now covers AI systems, models, datasets and prompts across ServiceNow and external platforms, with new connectors for Azure AI Foundry, Microsoft Agent 365 and Amazon Quick. Policies can automatically respond to a detected threat or block an agent, model or domain outright, including Explicit Block policies scoped to a department.
Containment is more practical now. A new AI Security Operator role can deactivate an agent without full AI Steward access, and the kill switch can reach agents running on Azure AI Foundry, not only on ServiceNow. The AI Gateway, generally available since September, sits between agents and MCP servers: it centralizes authentication, enforces policy, scans tools for poisoning before they are activated, and logs every tool call.
The piece to watch is AI Passport. Control Tower issues an attested record to each governed agent and MCP server that states what it may do; ServiceNow's October announcement with NVIDIA has OpenShell and BlueField DPUs reading that passport and enforcing revocations at the runtime and hardware layers. That enforcement is slated for a future release, so today it is a direction rather than a feature, but it tells you where the governance model is going.
Before you turn it on
04
Action Fabric was announced at Knowledge, and the MCP Server has been generally available and included in every Now Assist and AI Native SKU since then. Brazil is where the domain-specific servers land. The ITSM MCP Server exposes incident, change, request and on-call operations as tools an external agent can call: get and modify incidents with full business-rule execution, find similar incidents and knowledge semantically, run the change lifecycle including risk calculation and conflict checks, create requests with knowledge deflection and duplicate detection, and look up who is on call or request time off from a shift. ITOM, CMDB, SPM, Contract Management Pro and Service Mapping servers ship alongside it.
Access follows the authenticated user's ACLs and the domain logic runs server-side, so an agent gets the same answer a person would. The MCP Server Console gains a monitoring dashboard, OAuth client registration inside the wizard, a Quickstart server for summarizing incidents and cases, and tool annotations in the tools/list response so clients such as Claude can apply the right permission policy.
For a director, the point is simple. If someone in your organization is building an agent in another tool, this is the supported way for it to act on ServiceNow, with audit and metering, instead of a service account and the Table API.
Before you turn it on
05
The build experience is now one canvas: a visual node map of the whole agentic solution, a test conversation panel beside the configuration, and deployment from the same screen. You can add or remove agents and tools on the out-of-box agents without cloning them, which keeps you on the upgrade path.
AI Agent Advisor is the part we expect product owners to use most. It analyzes your actual case, incident and task data, proposes automation opportunities ranked by estimated annual hours and dollars saved, matches each to an existing prebuilt agent where one exists, and lets you create a custom agent from the opportunity in one step. The same opportunities surface in the new AI Admin Center, which also adds an AI readiness assessment, an upgrade readiness pre-check and an assist-consumption view by asset, user and department.
Governance tightened as well. New instances enforce deny-by-default ACLs on agentic record types, runaway agent triggers can be detected and disabled, and MCP servers get security controls defined before they are attached to an agent.
Before you turn it on
06
The L1 IT Service Desk AI Specialist is ServiceNow's first autonomous worker for ITSM. Assignment rules route incidents to it; it acts only on requests it is confident it can resolve, proposes a solution directly, and hands anything less certain to a human group. Brazil extends what it can do: it reads attachments such as screenshots, PDFs and error logs, it can resolve an incident by submitting the right catalog request on the user's behalf, and administrators can switch response templates between fixed text and a dynamic mode where the specialist writes work notes and customer comments from free-form instructions.
The monitoring is the part that will matter in a service review. A performance dashboard tracks every record the specialist touches, a Catalog tab shows drafted, submitted, abandoned and canceled requests by item, routing analytics show how many incidents were reassigned because a hand-off rule matched, and an AI Quality Assessment skill scores its work against a coaching rubric so a manager can compare quality over time.
The same foundation is meant to be deployed per team and customized: a configurable name and icon, a role and assignment group that bound what data it can see, and pre-built task configurations with your own routing behavior, templates and knowledge sources, no prompt engineering required.
Before you turn it on
07
Intelligent approvals takes a policy document, a PDF or a knowledge article, and turns it into approval logic. For each request it either auto-approves, auto-rejects, or steps aside and lets your existing approval workflow run. It sits in front of the hard-coded ask-for-approval step and only falls through when it cannot decide.
The home page for a policy shows a plain-language summary of what the AI understood, an editable trigger (table and condition), a pie chart of how many sample requests it could decide on its own, AI-generated test cases with reasoning, and a list of ambiguous wording in the document that is pushing decisions back to humans. In Brazil you can also build one from a knowledge article, with the policy automatically deactivated when the article is retired, or start one from a conversation with Otto.
Before you turn it on
08
Three things change managers have asked for in almost every ITSM engagement we have delivered arrive in Brazil. Change lockdown lets a change manager pause all or a subset of in-flight changes during a major incident, a financial close or any critical event; conflict detection flags any change scheduled inside the lockdown and puts it on hold until it is rescheduled. Scaled change generates and coordinates child change records for every affected configuration item, with scheduling intelligence, conflict detection and approvals keeping a large rollout on track.
Compliance dynamic schema stores risk and compliance attributes on a change as name-value pairs in a single field, so regulated teams can capture whether a change touches personal data, needs an audit trail or how long a rollback takes, without adding columns to the Change Request table. The base system ships example categories you can switch off and replace, and the values are readable from risk condition scripts and business rules.
On the AI side, four autonomous change agents are now available: creating a structured change from conversational input, suggesting the primary and affected CIs, drafting change plan fields during readiness, and picking the right template and model from historical data. State field policies for change models, which make fields mandatory or read-only per state, round it out.
Before you turn it on
09
Developer Sandboxes give each developer an isolated environment on top of a single sub-production instance, so parallel stories do not collide. Allocation is now instant because sandboxes are pre-created in a pool and refreshed every 24 hours, instead of being provisioned on demand in 15 to 60 minutes. And licensing is self-serve: sandbox packs are assigned from a new Developer Sandboxes tab in App Engine Management Center, without a support case, with a dedicated license admin role so you do not need to hand out full admin.
Every non-production instance now gets four free sandboxes, usable alongside purchased packs up to a cap of 34. Update set sources are configured automatically in both directions when a sandbox is created and removed when it is retired, and new REST APIs let you list, create, get and retire sandboxes from CI/CD or an MCP server.
Before you turn it on
10
Lux is ServiceNow's new UI framework, built on LIT, Base UI and Tailwind rather than the Angular 1 and Bootstrap 3 lineage that underpins Service Portal. Pages render on the server and stream to the browser, routing is file-based, widgets are Lit elements written once and reused across surfaces, and deployment is through the ServiceNow SDK. The practical consequence for hiring is that front-end developers from the wider market can be productive on it.
Brazil is where you start to see Lux in production surfaces: AI Control Tower, the EmployeeWorks web app, the redesigned ServiceNow Studio home page, and the new opt-in Flow Designer. That Flow Designer ships with the Flow Troubleshooting Agent, which investigates a flow that ran without errors but did the wrong thing, traces the data through each step, finds the business rule or configuration that conflicted, and hands back evidence with links to the records. It also lets you edit a step's inputs mid-execution in the debugger to reproduce an edge case.
Studio also absorbed the pro-code capabilities of the standalone IDE. New conversations start from the home page, you can right-click any record or configure menu and open it in Studio, source control is rebuilt with proper staging and sync, and you can view and edit the raw Fluent source of an app that has been converted. The Now SDK added five Fluent APIs (RestMessage, Alias, AliasTemplate, RetryPolicy, DataLookup), declarative record deletion and a now-sdk cicd command for pipelines.
Before you turn it on
Product names, availability and entitlements are ServiceNow’s and may change before general availability. Screenshots are ServiceNow’s and are reproduced for commentary. MadeLogic is a registered ServiceNow partner.