The BigQuery console is dead.
Agents run the queries for me now.
I can describe what I need, have an agent inspect the available tables, and ask it to write and execute the SQL. I don't have to spend the whole task moving between the query editor, table schemas, and previous queries.
That changes how much time I spend in the console. It also changes where I need to pay attention.
Getting a result is becoming easier
A lot of querying involves work around the actual question. Find the table. Check the column names. Work out how two datasets connect. Write the query, run it, and adjust it when something fails.
An agent can help with that sequence. It can also handle the next question without making me start again: split the result by region, compare it with the previous period, or investigate why one group looks different.
That is useful. It lets me spend more time on what I want to understand and less time manually carrying out each step.
But a query returning rows doesn't settle whether those rows answer the question.
The business definition still matters
Ask for last month's revenue and there are decisions to make before the SQL is useful.
Are we counting orders or payments? How are refunds handled? Which date defines the reporting period? Are internal accounts included?
An agent can generate a reasonable-looking query while making different assumptions from the people using the report. The result may look convincing enough that nobody questions it.
I still need to understand the data and the business question. If the definition is unclear, asking the agent to keep rewriting the SQL won't resolve it. Someone has to decide what the number is supposed to mean.
I still want to see what ran
When an agent produces a result, I want the SQL behind it. Which tables did it use? What did it join? What did it filter out? Did the join multiply rows before the total was calculated?
Those are familiar questions. Using an agent hasn't made them disappear.
I also want to know what the query will scan before it runs. BigQuery provides dry runs and query-cost controls; an agent should operate within those limits, using access appropriate to the task.
For an important result, I check a small part independently. Follow a few records through the calculation. Compare with a known total. Investigate a disagreement before copying the number into a report.
That takes judgement about the data, even when I didn't write the query myself.
Less console work, more questions about the answer
I expect to spend less time writing every query by hand as agents become part of the workflow.
The more interesting question is what happens to the work around those queries. Who defines the metric? Who recognises that the result is implausible? Who explains why it differs from the dashboard everyone already uses?
The agent can handle more of the querying. I still have to care about the answer.
Comments