Evals are the new product requirements document. Xavi Amatriain, Expedia Group's chief AI and data officer and former Google VP of AI and Compute Enablement, told the VB Transform 2026 audience in Menlo Park that all product thinking must move into evaluation suites before a single line of code is written. His argument is structural: encode what the product must do, including red teaming and security requirements, directly in the evals. With AI-generated code becoming the default, the eval becomes the only artifact that matters. VentureBeat's own Pulse survey of 157 enterprises puts the stakes in numbers: 66% already allow some production deployment without human review, yet only 5% fully trust the automated evals making that call. Half have shipped an agent that passed internal evals and then failed a real customer.
Amatriain runs governance at Expedia through three layers: principles first, then processes and tools to enforce them, then automation on top. He calls guardrails a necessary evil and argues that excessive rules corrupt feedback loops, biasing user signals and poisoning the model's learning. Expedia's alternative is risk-calibrated toll gates for agent releases, checkpoints that tie evaluation rounds, red teaming, and security review to each agent's actual risk level. Checks shift from recommended to required as stakes climb. His architecture is compositional, not monolithic: tools build into skills, skills into sub-agents, sub-agents into the full system. He held this view even at Google, where he rejected the AGI-as-singleton framing. The design detail on how Expedia unifies tone, context passing, and memory across that stack is worth reading in full.
The threat horizon is shifting. Amatriain told the audience that future attacks will come not just from humans but from other AI systems, making security a design principle that must be embedded as early as possible, not bolted on with guardrails after the fact. He described a feedback loop where production monitoring signals flow directly back into the eval suite, automating the entire correction cycle. VentureBeat's June Pulse survey of 107 enterprises shows how exposed companies already are: 54% have experienced an agent security incident or near-miss, and sandbox isolation, the primary post-breach damage limiter, drops from 35% adoption at smaller firms to just 20% at enterprises with more than 1,000 employees. Amatriain's toll-gate model is a direct bet against that trend. Whether calibrated governance can actually keep pace with autonomous agents is the open question this piece forces you to sit with.
[READ ORIGINAL →]