Your Teams Compensate for Missing Context. Agents Can't.

Scaled Agile has been releasing new guidance as part of the AI-Native SAFe series, and last week's instalment covered a topic that has been on my mind. The new articles on ART outcomes, Product Vision and Roadmap, and Intent, Specifications, and Context are all pointing at the same thing. An ART needs to build and maintain the product vision, ART outcomes, architectural constraints, and customer context that agents and teams reach for when they need to act. Andrew Sales put it plainly in session 2 last week: agents draw down on curated data. If it is not there or not current, no amount of prompting closes the gap.

This is not a new problem. It is a more visible one.

Those of you who have worked with us will know our feature discovery pattern. This is our approach to ART backlog refinement, designed to ensure ready, prioritised features for PI Planning. Every team on the ART reserves a percentage of their capacity every iteration to work on feature definition (discovery) for features targeted for the next PI. Teams use this reserved capacity to work with Product Management, the System Architect, and subject matter experts to align on the benefit, acceptance criteria and high-level solution for candidate features. This works best when Product Management has a vision and roadmap that is supported by a clearly articulated architectural runway developed by the System Architect.

Organisations that have this pattern nailed will find the new guidance less of a wake-up call and more an expanded vocabulary for something they were already building toward: a curated data stack. For everyone else, the urgency just arrived before the foundation did.

Delivery isn't the (biggest) problem anymore

Recently, I have been invited behind the curtains at a number of organisations, some past customers, some entirely new, and what I am seeing is a new pressure that did not exist two years ago. For the ARTs that haven't yet built a strong feature discovery practice, the squeeze is real: the pace of AI-assisted delivery has outrun the ability of Product Management to feed the ART with a well-defined, strategically coherent backlog.

Without a product vision, an outcome-driven roadmap, and an architectural runway that looks out further than the next PI, the ART runs out of material. Or worse, the ART's attention is diverted to random work disconnected from the product's strategic direction, in the name of "resource efficiency", the efficiency paradox Lean has been warning us about for years.

Feed the agent, or it feeds itself

An agent doesn't know what it doesn't have. It can only work with what is there. The product vision, the ART outcomes, the architectural constraints, the customer and market environment: if these things do not exist, or exist somewhere but have not been maintained, the agent reaches for them and finds nothing useful. So it does what any self-respecting autonomous agent would do. It finds the open door, sprints straight down the hallucination hallway, and keeps going. No one told it to stop.

That is not a technical failure. That is the predictable output of an underfed system. And the fix is not better prompting; it is the curated data that should already be there.

ARTs that were already struggling to get a coherent vision and roadmap in place before AI came along have not suddenly solved that problem. They have supercharged the gap, not closed it. The gap that was manageable when humans were compensating becomes something else entirely when agents are amplifying.

The guidance arrived at the right moment. Use it.

Honestly, I am seeing more pain than success out there. The new articles give Product Management and the System Architect something concrete to work with, not just a diagnosis of what is missing, but a framework for building it. We consider feature discovery a "turn up the good" pattern for ART backlog refinement. It is time to turn up the good again.

The agents are already here. Are you ready to feed them?