Top features in Brazil release Part 1

Brazil Release for Developers (Part 1): From Requirements to Running Code with Autonomous Engineer and Build Agent

With Australia, developers learned to work alongside AI. With Brazil, the role changes again: developers increasingly supervise it. AI now covers more than generating components. It reaches into planning, test generation, troubleshooting and deployment packaging. The developer’s job moves toward approving plans, reviewing output, and owning the decisions that matter: architecture, data model, security and business outcome.

Brazil Early Availability arrived on 24 September 2026, and General Availability is expected in November 2026.

This series is split by lifecycle stage:

Part 1 (this post): Plan and Build. Covers Autonomous Engineer, Build Agent, and the code-first toolchain.
Part 2: Isolate, Debug and Govern. Covers Developer Sandboxes, the Flow Troubleshooting Agent, playbooks and testing, and MCP governance.

1. Autonomous Engineer: specification-driven delivery

What it does

Autonomous Engineer is an agentic worker that runs on top of Build Agent and takes a written specification as its starting point. It works in two phases.

Plan:

You supply requirements, either as a prompt or an uploaded document.
It interviews you to resolve ambiguities.
It queries your instance for existing artefacts it can reuse.
It produces an implementation plan of work items, each with acceptance and test criteria.

Build (only after you approve the plan):

A background agent picks up each approved work item, and the items are built in parallel.
It generates and runs ATF tests and attempts to fix failures.
Anything it cannot resolve is surfaced for human review.
The finished work is exported as an update set for UAT or production.

Agent packs

When a plan touches a ServiceNow product (not just a custom app), Autonomous Engineer loads an agent pack for that product. A pack bundles product-specific knowledge, tools and skills, plus Test Agent capability for that product. If a plan mixes product work and custom development, each pack applies only to the work items that target its product.

Licensing: check this first

Autonomous Engineer is a separate ServiceNow Store app. It requires the AI Workflow Factory SKU and a Brazil EA (or later) instance. A Build Agent entitlement on its own is not enough.

Architect’s take

Treat the plan-approval step as a design review, not a click-through. Before you approve, check:

Does it extend existing tables, or create new ones you don’t need?
Is the application scope correct?
Does the role and ACL model follow least privilege?
Are naming conventions followed?

Autonomous Engineer fits work whose scope is clear enough to write down. It is the wrong tool for discovery-phase problems where the requirement itself is still being shaped.

2. Build Agent: one conversation from build to package

Build Agent is the conversational engine underneath all of this. In a single chat it can:

create and modify full-stack applications, both UI and server-side;
explain and document existing code;
run quality checks;
generate and execute ATF suites, and report test errors back in the chat;
build custom agents, skills and agentic workflows;
package changes for deployment.

Checkpoints let you roll back at any point in a session.

What changes in Brazil

On by default. Build Agent (Trial) is active by default, with no Store install or admin request needed. It does require ServiceNow Otto for Creator.
Studio and IDE converged. ServiceNow Studio has merged with ServiceNow IDE. New conversations start from the Studio home page, and you can toggle into Autonomous Engineer from there.
“Open in Studio” shortcuts. These now appear on any sys_metadata form, on the script editor toolbar, and in Configure menus.
Auto-upgrades. Qualifying instances (Brazil P0+, Australia P5+, Zurich P12+) now receive new Build Agent versions automatically. Upgrades run as a scheduled job, so a new version can take up to a day to land.
Wider MCP support. Box has been added as a connectable MCP server, and a settings pane shows which tools each connected server exposes. Some of this arrived in late Australia patches, so check the minimum patch level for your target instance.

Gotchas to plan for

Admin role only. Build Agent currently requires the admin role. If you run a delegated-development model on least privilege, decide how you will handle this before standardising on Build Agent.
No domain separation. Build Agent will not configure domain separation itself. That stays a deliberate, human-designed decision.
Restricted environments. Availability is limited in some environments: FedRAMP, NSC DoD IL5, Australia IRAP-Protected, self-hosted, and certain regulated markets (see KB0743854 and KB2593939). Confirm availability before you promise it to a customer.

Example use case

You inherit a heavily customised app with thin documentation. You can use Build Agent to:

Map its tables, business rules, flows and roles.
Make the change.
Generate and run an ATF suite.
Review any failures.
Regenerate the documentation.

The time saved comes mostly from understanding and safely changing the app, not from writing new code faster.

3. Code-first is converging with agentic

ServiceNow SDK. Supports local, source-controlled development in VS Code Desktop. Prerequisites are Node.js 20.18.0+ and npm 8.19.3+, and the SDK works with instances from Washington DC onward.
ServiceNow Fluent. A TypeScript-based declarative language for metadata such as tables, roles, ACLs, business rules and ATF tests. Some metadata types are still represented as XML.
Why this matters: Build Agent generates metadata that Fluent supports. That means AI-generated work and hand-written code-first work now share one representation, which is good news for Git-based workflows.
Lux Lab for VS Code. Still useful for scaffolding experiences, pages and widgets, and for previewing them locally.

Closing thought

Brazil takes away much of the repetitive effort between a requirement and a deployable update set. It does not move accountability. The teams that get the most from it will be the ones that decide in advance where human approval is mandatory.

Coming in Part 2: isolated sandboxes, AI-guided flow debugging, playbook permissions, broader test coverage, and governing MCP connections.