The data layer retools for agents: adaptive compute, memory, and agent-scale git

Compute that sizes itself

Snowflake's Adaptive Compute, generally available since June 16, 2026, dissolves fixed warehouse sizes from XS to XL into one shared compute pool that sizes every query, the company explained. Instead of picking a t-shirt size, users manage two dials — MAX_QUERY_PERFORMANCE_LEVEL and QUERY_THROUGHPUT_MULTIPLIER — and pay per query based on the compute and software resources each query actually used, rather than paying while a cluster runs. The design target is explicit: unpredictable, agent-generated SQL workloads. Snowflake reports 1.2x better price-performance than Gen2 on its TPC-DS 10TB concurrency benchmark. There are sharp edges: resource monitors disable an Adaptive Warehouse rather than suspending it, and per-query attribution views have gaps teams should check before migrating cost dashboards. m:…7186

Discovery and memory

For finding data, Snowflake launched CoCo, an agentic discovery tool for its public marketplace now in public preview as of October 9, 2026. Users describe data needs in natural language instead of guessing keywords, and can install, trial, or request sales directly from the conversation — with actions that are role-aware and surface only when the user's role has the required privileges. The honest limits are stated up front: CoCo does not yet perform structured cross-listing comparison on arbitrary dimensions and does not return pricing information inline, so specification verification stays manual. m:…6153

For agent accuracy, Snowflake's Haebichan Jung published an RLHF-inspired method for long-term memory that updates the context an agent retrieves at inference time instead of retraining model weights. The five-step loop — capture feedback, diagnose with AI_COMPLETE, review, store approved lessons as rows in a Snowflake table indexed by Cortex Search, retrieve — took a retail sales agent from 48 percent to 100 percent accuracy on 33 new test questions, with no latency impact, the company said. Human feedback acts as the reward model, entering through retrieved context rather than gradient descent. The code is open source. m:…6266

Snowflake also integrated NVIDIA Kumo into Snowflake ML, letting data science teams train graph neural networks directly on relational data — learning from tables, keys, and temporal patterns rather than flattened wide tables — with data never leaving the customer's account. Early users include Affirm, National Oilwell Varco, VideoAmp, and Yieldmo, per the announcement. NVIDIA separately announced the Agent Knowledge Foundation, with open models and libraries to help enterprise agents retrieve evidence, apply consistent business meaning, and predict outcomes. m:…0345

Git at agent scale

GitHub is rebuilding its Git infrastructure for the same reason: agents. The company is migrating to Azure, saying the timing and shape of the move were driven by the agentic shift — rapidly growing traffic, a hard ceiling on data-center power capacity, and physical supply-chain constraints. Git traffic has grown exponentially over the past year, and 99.99999 percent of repositories now live in both GitHub data centers and Azure, with Azure serving about half of Git traffic. The architectural changes are substantial: per-reference serialization removes the repository-wide write bottleneck (3-4x higher throughput at lower latency in internal tests), and decoupling storage from compute lets read replicas scale on demand, warming caches from Azure Blob Storage. m:…1835

What it means

This is my read, not a reported fact: the throughline is infrastructure being re-architected from human-paced to agent-paced usage — bursty generated SQL, parallel writes, persistent memory. Developers building agents inherit new primitives: self-sizing compute, retrievable long-term memory, elastic version control. The cost-model shift matters as much as the capability shift: per-query billing plus agent-generated query volume is a new metering discipline, and teams should instrument it before agents scale, not after the first surprise bill. m:…7186 m:…1835