System Architects used to design architecture up front. In an AI-Native ART, they build guardrails directly into the workflows teams use every day.
What the role was
In core SAFe, the System Architect defines and communicates the shared technical vision for the ART. They guide the architectural runway, shape enablers and set the non-functional requirements the system must meet. Much of that guidance lived in designs, standards and reviews.
That worked when people did the building. People read the standard, remembered it, and applied judgement when it didn’t quite fit.
What changed
AI agents now do a growing share of the work, and they don’t behave like people. The framework is direct about this: given a poorly defined process, an agent improvises — and improvisation at machine speed produces inconsistency, obscures accountability and scales mistakes.
An architecture that only exists in a document is exactly the kind of undefined process an agent will improvise around. So the architecture has to move into the work itself.
What guardrails in the workflow means
The AI Innovation Pipeline’s Embedded Policies component is where this lives. Policies fall into three categories:
- The system qualities a solution or workflow must exhibit, such as performance and security
- The standards it must meet, from encryption protocols to regulatory obligations
- The checks it must satisfy, such as a definition of done or an approval gate
Rather than inspecting finished work at a stage gate, compliance is built into every step. Once a policy is stated and made machine-readable, it applies to every AI-empowered workflow. The framework’s point is that guardrails built this way are the condition for speed, not a brake on it.
Two design choices the architect now owns
Where a human sits. The framework distinguishes human-in-the-loop, where work pauses for approval, from human-on-the-loop, where work runs within bounds while a person supervises. The rule is proportion: the more consequential or irreversible a decision, the more it warrants a human in the loop.
How much authority an agent gets. Agent authority is granted, bounded and earned. Each agent works under an identity that scopes what it may do, which tools it may use, how much it may spend and when its access ends. It starts narrow. Reliable performance expands its scope; failure narrows or revokes it.
Neither of those is a one-time design decision. Both have to be maintained as the system and the agents change.
What doesn’t change
The architect still owns the technical vision and the runway. What changes is how it reaches the work. A standard that teams must remember becomes a policy the workflow enforces — for people and agents alike.
Three questions for your System Architects
1. Which of our architectural standards are enforced inside workflows, and which depend on someone reading a document?
2. For each AI agent we run, what is its scope — and who decides when it expands?
3. Where do we have a human in the loop, and is that proportional to how reversible the decision is?
This is one of four pieces on how ART leadership changes in AI-Native SAFe. Read the overview of all four roles.
The world has changed. The leaders who notice will be the ones the next decade is built around. I write on Thursdays at agile-agilist.com.
Sources
- Scaled Agile, Inc. — AI-Native SAFe framework guidance: AI Innovation Pipeline (Embedded Policies)
- Scaled Agile, Inc. — SAFe 6.0: System Architect
