Source: devblogs.microsoft.com/foundry/when-one-pass-isnt-enough-agentic-mode-in-azure-content-understanding
Microsoft has added an agentic mode to Azure Content Understanding, now in preview, for document tasks that a single extraction pass handles poorly. Standard mode remains the default and reads a document once to fill schema fields. Agentic mode runs a reasoning loop with tools for navigating text and tables, inspecting visual evidence, calculating and validating results.
In a Treasury bulletin example, the answer requires selecting values from a table that spans two pages, excluding some entries, normalizing notation and then calculating a total of 180,681 million dollars. In a network log spanning 19 pages, the tool consolidates rows and counts or sums events matching given conditions, rather than inferring a count from event IDs that contain gaps.
Microsoft says agentic mode uses more model tokens and typically takes longer on the same document, so it recommends comparing both modes on representative files. To try it in Content Understanding Studio, users select the 2026-06-01 (Preview) API version and the Agentic Extraction workflow kind. In code, the analyzer config adds "workflow": "agentic" with API version 2026-06-01-preview.
Why it matters if you build agents
Developers building document agents can route only the hardest documents to agentic mode and keep standard mode for the rest. Before switching, they should compare both modes on their own files, since agentic runs cost more tokens and time. The preview requires the 2026-06-01-preview API version and a supported completion model in the analyzer definition.
Get stories like this every weekday
One short email with the day's agent tooling news.