The Sandbox is the Product
Something is changing in product management that nobody has the vocabulary for yet.
For two decades, the PM craft was built around a deterministic system. The PM specified the artifact, engineering built it, the artifact behaved the way the spec said it would, and the metrics measured whether users used it. The PRD was the central document. The roadmap was the central plan. The success metric was usage.
That model is breaking. Agentic systems do not behave the way the spec says they will. They behave the way the interaction among the spec, the reward signal, the feedback geometry, the recovery mechanics, and the iteration budget pushes them. The PM no longer specifies the artifact. The PM designs the conditions under which the artifact emerges, observes the emergent behavior, and tunes the conditions until the emergent behavior matches the business need.
The conditions, in their concrete form, are a sandbox. The sandbox is the new product. Not the agents inside it — those are commodities, replaceable, increasingly capable, increasingly cheap. The sandbox is what determines whether the agents produce compounding value or an avalanche of plausible-looking nonsense.
Most PMs operating agentic products in 2026 are doing sandbox design without knowing it. They have a system prompt that nudges agent behavior — that is the implicit reward. They have a set of tools the agent can call — that is the implicit action space. They have logs nobody reads — that is the implicit observability layer. The sandbox exists. Nobody has named it. Nobody is governing it. The agents inside are reward-hacking quietly, the metrics look fine on the surface, and the team has no vocabulary to describe what is going wrong.
This piece gives that vocabulary. It is structured as a working manual: the five components of a sandbox, the deepest sub-craft (reward design), the most under-built layer (observability), the maturity arc that runs from v0 to v3+, the economics of when sandbox investment pays off, and a worked example of building one from a blank page to running v1.
If you are a PM, an agent-product engineer, or a product leader trying to govern teams shipping agentic systems, this is the field manual.
For Exec Members, The Full BE Agent Harness Is Now Available!
