Two agents edit different files. Git merges clean. The program is broken. No conflict was thrown, because the collision was in meaning, not in text — and business decisions are almost all semantic. The fraud counter and the account balance don't touch the same row, so there's nothing to lock and nothing to catch. They're simply incoherent, and the agent approves on the incoherence.

This failure is structural. For decades, applications got machine speed with thin data, and humans got rich context with no urgency — staleness was free because the human deliberating for three days was the bottleneck. An agent is both consumers in one body: an analyst's context at an application's latency. The bottleneck moved from deciding to observing, and nothing was built for the new position.

Correct context has to be three things at once: shared (concurrent agents on one coherent snapshot), live (the world at the moment of decision), and semantic (derived signals and retrieval by meaning). Every system in a typical stack gives you some. None gives all three — and composing more makes it worse, since each addition is another clock answering about a different instant. Five systems return five honest answers about five different moments, and nothing throws.

I'll show the failure concretely, then the architectural requirements for closing it: incremental materialization, hybrid row/columnar storage, and unified retrieval under one snapshot. Including the honest limits: ingested context converges sub-second, it isn't instantaneous.a