Your agent's answers are only as honest as its context

Wrong context is worse than none

An agent that reads the wrong data does not just give a wrong answer. It acts on it. DataHub's argument this week is that AI-ready data lineage is not a map of where data lives — it is an accurate, automated, column-level map of how data flows across every system in a stack, and accuracy on ambiguous SQL is the property that decides whether an agent can be trusted with it. For an agent, the post says, wrong lineage is worse than no lineage, because the agent acts on it. Accuracy and context are the two properties most evaluations of lineage tools miss. The example given is Miro, whose text-to-SQL accuracy improved from under 40 percent to over 90 percent after implementing a ranking framework for metadata signals such as schema ranking and usage quality gates. The model did not get smarter. The context got honest. m:…3018

Delphina, a startup building what it calls an enterprise context factory, makes the same point from the agent's side. Coding agents frequently make mistakes when working with enterprise data because they lack institutional business context, the company says, and an accuracy rate of 80 percent is insufficient for AI agents used in important business decision-making. Its Context Factory uses agents to discover, organise and maintain institutional knowledge from documentation, dashboards and historical SQL queries, continuously reconciling them so the agent's answers can be validated. The founders previously built data and AI systems at Uber, GoPuff and Tecton. The pitch is not a better model. It is a maintained memory of what the business means. m:…2343

Governance becomes architecture

Snowflake's contribution this week is a reference architecture that turns governance into structure. Its Operational Ontology for AI agents connects business definitions to operational behaviour through six capabilities: governed meaning, operational state, operational semantics, decision services, AI reasoning, and governed execution and verification. The decision lifecycle follows six steps — Observe, Understand, Evaluate, Decide, Execute, and Verify — and the key principle is to separate meaning from behaviour while connecting them through explicit contracts. Simulation services, implemented with Snowpark stored procedures, let agents evaluate whether a proposed plan is feasible before executing it. The agent is allowed to act, but only inside a contract that says what the words mean and what the action changes. m:…3646

The catalog underneath has to keep up, and DataHub's second post this week describes the shift from static portals to context platforms: automated catalogs that use AI and event-driven metadata to discover, document, classify and sync data assets continuously, delivering governed semantic context to people and AI agents wherever they work. Pinterest eliminated about 70 percent of manual documentation work by using AI-generated documentation and lineage-based propagation, the post says, and 63 percent of IT and data leaders placed their organisations at the third or fourth stage of data catalog evolution in the 2026 State of Context Management Report. Human review remains necessary, the post adds, to define standards and resolve conflicting definitions. The machine writes the catalog. The human still decides what the words mean. m:…3766

The factuality trade-off

Two research papers this week quantify the cost of getting context right. One applies conformal factuality to agentic retrieval, where the agent must decide when to stop retrieving. Stricter conformal targets are consistently associated with higher observed response-level factual support, the authors find — but increasing them reduces claim retention and non-empty response coverage. An agent that is forced to be right answers less often. The trade-off varies significantly across models and across datasets, which means there is no single setting to copy. Every deployment has to choose its own point on the curve, and say so. m:…5106

The second paper looks at retail, where an agent must chain evidence across heterogeneous data sources. Multi-source evidence chaining is the dominant bottleneck for LLM agents in retail workflows across providers and model scales, the authors find, and reliable retail agents require explicit cross-source orchestration rather than scaling model parameters alone. Their RetailRAG framework, built on auto-schema briefing, source-routed plan chaining and diagnostic plan revision, outperforms existing agentic RAG baselines across all evaluation axes. The week's pattern is consistent across vendors and preprints alike: the model is no longer the scarce part. The scarce part is context you can prove — lineage that is accurate, meaning that is governed, and answers that know when to stop. m:…5105