mode set to agentic. It parses in several steps and spends the most work on dense tables, so it is the tier to reach for when the Standard output on a complex layout is not good enough.
The document
INVOICE
→
What comes back
p1_b0100
p1_b586
p1_b631
document
NodesCredits100 per pageYou get
The graph
effort is an Agentic-only knob:
prompt_hint also applies, for example "the second column is a date". figure_enhancement is ignored on this tier.
Run and read
Agentic runs are slower than Standard: 37 s for a one-page invoice against 28 s, and minutes for a long document. Givewait() a bigger budget.
What comes back
The Agentic tier on a one-page invoice, trimmed. The shape is the same as every other tier; only the work behind it differs.Response
classifications, splits, extractions and edits all come back empty, because only Parse ran. Runs and results has the full envelope, section by section.
Reading the confidence
blocks[].parse_confidenceis the number to act on. It is 0 to 100 per block. In the sample above the line-item table scored 86 and the totals block, which the tier read as a second table, scored 31. That is the block to send to review.parse.parse_confidenceis the document roll-up, weighted by the characters in each block. A long low-scoring table pulls it down more than a short heading pulls it up.- A block that no model scored comes back with
parse_confidence: null. Treatnullas “not scored”, not as zero.
When it goes wrong
Every document type is on Agentic. Agentic costs 100 credits a page against 25 on Standard, four times the price for the same graph. Try Standard first, and move a document type up only when its tables come back wrong. One table came back as two. In the sample run the tier read the totals as a second table and scored it 31, next to 86 for the line items. That is the block to send to review, andblocks[].parse_confidence is what tells you so. Read the block’s rows for the cells as a grid rather than parsing its <table> HTML.
The document score disagrees with the blocks. parse.parse_confidence is a roll-up weighted by the characters in each block, so one long low-scoring table drags it down further than a short clean heading lifts it. Judge a run on the per-block numbers. A block that no model scored comes back with parse_confidence: null, which means “not scored”, not zero, so test for null before you compare.
The run times out. Agentic takes about 37 s on a one-page invoice against 28 s on Standard, and minutes on a long document. Raise the budget you hand to wait(): timeout=600 in Python, { timeoutMs: 600_000 } in TypeScript.
A dense table is still wrong at effort: "mid". Move to accurate for scanned, low-contrast or very wide tables. Drop to low on clean digital documents where you want the multi-step pass at a lower price. prompt_hint takes a sentence about the page, for example "the second column is a date". figure_enhancement is ignored on this tier, so it will not help here.
To turn the same tables into typed fields, add an Extract node after the Parse node. Bank statement processing does exactly that.
Next steps
Parse node
Every knob on the node
Parse-only workflow
The same graph on Flash, Fast and Standard
Get run
The full run envelope
Coding assistant
The anyformat skill for Claude Code and other agents

