1. What are the benefits of standardization? 2. What open standards do you think are missing to enable the new data & AI workloads we are seeing today?

1. GPUs get replaced, moved between servers and split into partitions. Do you key their error history on the physical card, the server slot or the device the scheduler sees? 2. A GPU can be split into MIG slices or AMD partitions, but power is reported only for the whole card. How do you measure each team's usage, and does your clawback logic go by the slice reserved or the work done? 3. Goodput needs three levels of data: job runs with start and end times, failures and checkpoints. Which one does your platform land today, and who owns the number and the data quality?

1. When several sources can each answer the same question with different numbers, where should the decision about which one to trust be made: in the agent's reasoning, in the routing layer, or in the data itself? 2. When two sources differ for legitimate reasons, whose job is it to carry that distinction, and how much of it should the agent ever see? 3. What should an AI have to show before you trust its answer, and how would you know your evaluation is actually testing that?

1. With v3 features like deletion vectors, row lineage, and the variant type now shipping, how long will engines and catalogs lag behind, and how should teams decide when to upgrade a table's format version? 2. Single-file commits and the adaptive metadata tree are the headline v4 proposals. Do they finally make Iceberg a good fit for high-frequency streaming and small writes, or does compaction stay a permanent tax? 3. Is the REST catalog becoming the real interoperability standard, more than the table spec itself? What still needs to be standardized: credential vending, fine-grained access control, or events?

1. When an agent acts on an employee’s behalf, whose access should it get: the employee’s, its own, or only what both are allowed? 2. Who should own AI data policy: the data team, security, or the teams building with AI? 3. Should AI spend be a live control that can throttle a team or agent mid-task, or does that make finance your AI policy owner?

1. What evidence actually gates your decision to move an AI feature from pilot to production, and who signs off on it? 2. When an agent gives a bad answer, what evidence do you need to determine why it failed and which part of the system actually caused the failure? 3. AI systems change faster than our evaluation processes. How do you decide when you have enough evidence to ship, rather than continuing to evaluate indefinitely?

1. What has paid off for your customers? Where have faster responses, fresher data, or more flexible queries made a clear difference to customers, and where did you build for needs they don't value? 2. What did you trade away as you scaled? When speed, flexibility, cost, and engineering simplicity pulled in different directions, what did you give up, and what wouldn't you give up? 3. What changes when agents query the data? Which requirements are new, which extend existing customer-facing workloads, and how much should you build before agent usage is proven?

1. When an agent-driven data workflow pauses or fails, what should it preserve, and what should it revalidate before resuming or retrying? 2. How should teams combine human approval and automated evaluations to decide which agent-driven data changes can proceed to production? 3. How should teams design recovery for agent-driven changes that are partially complete, affect large datasets, or have already reached downstream systems?

1. What needs to be done differently when preparing a data product for AI agents - what needs to be decided upfront, and what can be introduced or learned incrementally? 2. How should ownership and responsibilities be defined across data producers, data/AI platform, agent builders and business domain owners? 3. What evidence would convince you that a data product is ready for agent use, and when should an agent defer to a human?