I lead data at a fintech and run AI agents on our real stack (Snowflake, dbt, n8n). Writing down what breaks. Building data-trust-skills (open source).
For official KPIs, a semantic layer your agent can silently bypass is just a suggestion.
I'd keep two routes:
Reporting → approved metrics.
Exploration → scoped SQL, clearly labeled.
Promote repeat questions into metrics. Don't let exploration impersonate reporting.
my favourite agent move is when you ask why the number changed since yesterday and it gives you three confident, nicely formatted reasons. none of them is 'the load hadn't finished'
@isidoremiller fair for dashboards. for agents it's actually binding though. an analyst shrugs at three definitions of revenue, an agent just confidently picks one
@morganlinton per-task effort is the part i care about most. don't need max reasoning to rename a variable, very much need it when it's touching a join
Nvidia just fell below $4 trillion.
Nothing about the chips changed this week.
Nothing about the roadmap.
The only thing that changed: the guy in the leather jacket is gone.
Key-man risk was the most underpriced line in the entire AI trade. Not anymore.