The most important sentence in today's signals comes from Berkeley: frontier inference has fallen from $30 per million tokens to under $0.10 in roughly two years (BAIR). Read that as a verdict. The model is no longer the moat. If your 2026 AI strategy still opens with a model-selection debate, you are optimizing the one variable the market has already commoditized.
Watch where the value went. It did not evaporate—it migrated to the Decision Surfaces and the Identity Control Surface, the places where agents touch data, money, and action. Today's cluster makes that migration concrete. Mastercard is now writing the agent authentication spec across three converging protocols and has shipped Agent Pay for sub-second machine-to-machine settlement. Okta is hardening vendor-neutral identity governance for non-human agents. When the payment rails and the identity layer standardize this fast, agent governance stops being a whitepaper concern and becomes a gate. Enterprises without a non-human identity inventory will simply be locked out of the commerce rails their competitors ride.
And most are not ready. KTSL and BMC Helix found 88% of large UK enterprises deploying agents, but 25% reporting ROI shortfall—driven by governance gaps, data quality, and skills (ITSM.tools). This is the Compiled Corporation failing in the field: firms automating decision-making without the scaffolding to make those decisions trustworthy are automating noise. The index reflects the same fault line. Governance & Ethics is our strongest organizational mover at 75, and Agent-Ready Infrastructure ticked to 49—but Product stalled flat at 57. Product-layer AI is shipping ahead of the organizational and identity foundations that let it compound. That gap between deployment velocity and governance maturity is the 25%.
Cisco shows the other path. Rolling agents to 90,000 employees framed explicitly as change management, not IT rollout, with on-prem data control, is a textbook Janus Brands move—an identity shift announced as one. That is the benchmark large enterprises will be measured against this planning cycle.
The strategic instruction is singular: stop shopping for intelligence and start building the identity architecture that governs it. Inventory your non-human identities before Mastercard's rails demand you have. Wire treasury and procurement for agent access before your agents route around them. The firms that win the next 18 months are not the ones with the best model—they are the ones who control the pipes.
Watch this: Microsoft Research's SkillOpt converts manual agent-skill editing into a trainable process without touching base weights—directly targeting the brittleness behind that 25% failure rate. Track for productization in Azure AI Studio or Copilot Studio over the next 60 days. If it ships, agent behavior becomes reliable enough to actually govern.
WatchSkillOpt (Microsoft Research) converts manual agent skill editing into a trainable process without modifying base model weights. If this technique reaches production tooling, it directly addresses the brittleness problem that is driving the 25% agent ROI failure rate—making agent behavior reliable enough to govern. Track for productization signals in Azure AI Studio or Copilot Studio over the next 60 days.