Not a list of dated SKUs.
The trust curve is the actual roadmap.
Every percentage point of verified truth added to the model raises the ceiling on how much an agent can do without a human checking. The features are what you earn as the curve climbs.
Agents graduate as the model earns trust.
Trust is a number per edge, aggregated into a number per region of the model. Action class is gated on the confidence floor of the region the agent is reasoning over.
The agent answers what is happening with cited evidence. The human acts. The five agents live here today: they surface the verdict and refuse to invent one when the evidence is not there.
The agent answers and proposes a discrete action with a confidence. The human approves. The action runs. Reversible actions earn this tier first.
For narrow, high-confidence cases, the agent acts on its own. The human reads the receipt. The classes of action expand only as the corresponding confidence floor of the model rises.
The verified surface of the model is wide enough that operational labor moves to the agents by default. Humans set policy, write the model's rules, own the exceptions.
What the model makes buildable.
Five agents live today. More sit at the edge of buildable. Not a price list.
A corpus of system shapes.
Once many companies run the same model, you hold a corpus of system shapes. What healthy looks like. What rot looks like. The patterns that precede outages.
Cycles of half-removed code, leftover consumers, drift between declared and observed. Recognised across the corpus, surfaced before they bite.
Specific shape changes that historically precede specific failure modes. Anonymous, aggregated, alertable at the model layer.
Every question an architect, SRE, or agent has ever asked the model. The good queries get reused across customers without leaking their data.
This is the moat the single-tenant AI-SRE and the manually maintained catalog structurally cannot reach. We are walking toward it slowly because the substrate has to be honest first.