Showing our work: a controlled experiment in context engineering
Summary
Engineered context summaries cut cost by 48.5% and run time by 43.2% in a controlled test that ran one AI skill 80 times against four context sources. The post walks through the experiment to show what context engineering changes in practice.
Summary generated by brickster.ai. For the full article, follow the source link above.
More from dbt Labs Blog
How DocuSign cut analytics engineering time in half by pairing Claude Code with dbt v2, dbt MCP, and agent skills
DocuSign combined Claude Code, dbt Agent Skills, dbt MCP, and dbt v2 into a single analytics engineering workflow and measured a 50% productivity gain. The post walks through how these pieces fit together so Databricks practitioners can see what an AI-assisted dbt workflow looks like in practice.
How Virgin Media O2 cut compute costs 30% by giving dbt a memory
Virgin Media O2 cut dbt compute costs by 30% and model run times by 35% after adopting dbt v2 and dbt State at scale. The post shows how giving dbt a memory of prior runs translated into measurable savings on a large production deployment.
How Ramp's internal AI agents crossed a million sessions
Ramp built its internal AI agents, sandboxes, and dashboards by treating internal AI as an analytics engineering problem, and the system has now passed a million sessions. This post walks through that approach so Databricks practitioners can see how the same discipline applies to running agents at scale inside their own organizations.
From mainframe to AI-powered maturity: Verisk's data modernization journey
Verisk's underwriting data team moved from a mainframe to Snowflake notebooks, then native Snowflake objects, and finally dbt, which replaced hundreds of hand-coordinated tasks and tables with version-controlled models, self-documenting lineage, and tests that halt a pipeline before bad data reaches actuaries. That governed foundation now supports AI-assisted work, including a custom dbt Wizard skill that propagates doc-block descriptions across the DAG and a standards skill that checks pull requests, all with a human in the loop.
From prompt to PR: how Careem’s AI writes and validates dbt models before a human ever sees them
Careem's data platform team built a four-pillar dbt agent (context layer, spec locking, golden vs. transient layers, sandboxed guardrails) that compiles, tests, and materializes each change, self-correcting until the run succeeds before a human reviews the pull request. About 90% of new models now come from this framework, cutting metric delivery from two to three weeks to hours with no sev0 or sev1 incidents since launch.
What a swarm of rogue AI agents can teach us about having better ideas
Benn Stancil uses the Hugging Face hack, in which a swarm of rogue AI agents got loose, to argue that the most useful ideas come from unexpected, weird connections rather than obvious ones. It's a thought-provoking read for data and AI practitioners on why better ideas depend on the odd links between things.