Beyond Prompt Engineering: Ambient Context Is the Next Axis

AI capability has advanced in three waves. First, prompt engineering — getting better at telling the model what to do. Then RAG — wiring the model to documents it could pull from when its memory wasn't enough. Both made models more capable. Neither solved the harder problem, which is that the model still doesn't know what the user is doing right now.

Ambient Context Engineering — ACE — is the third wave. It's the layer that tells the model what's currently on the screen, what the user just did, what application is open, what record they're staring at. Not the prompt. Not the indexed document. The live state of the work.

The distinction matters. RAG can pull a relevant policy document for the user's question, but only after the user has phrased the question. ACE can tell the agent the user is on the policy-violation screen for customer 4471 before the user types anything. One acts on a question; the other acts on a situation.

ACE is harder to ship than RAG, which is why most of the field hasn't gotten there. It requires a substrate on the desktop, a model of what work even is in this organization, and a multi-year record of activity to learn the patterns from. The capability is gated by infrastructure, not by language.

OLi is built ambient-by-default. The prompt becomes the exception — the rare case when the agent needs the user to disambiguate — instead of the every-case toll booth between the user and any useful output.