Separate deterministic checks on sales CSVs from summarizing campaign notes, then combine them into a human-reviewed draft with definitions, exceptions, and open questions.
Use formulas to check orders, amounts, and currencies, then have a model organize campaign notes into a sourced draft. Review figures, denominators, and exceptions. Sum amounts only when currency and definitions match; list missing, duplicate, or conflicting records and pause affected conclusions.
CSV columns, amount totals, currencies, and duplicate-order checks follow explicit rules and are usually better handled by spreadsheet formulas or code. Campaign explanations, actions, and context require reading text, so a model can draft a source-linked summary. Keep the two jobs separate so the model does not recalculate business figures.
A weekly flow can be: receive approved files → check columns, dates, duplicate orders, and currencies → total figures with matching definitions → extract campaign updates from notes → create a draft with open questions → have an owner review before sharing.
Before adding amounts, confirm the currency and whether refunds, tax, shipping, or order status are included. Keep currencies separate; if conversion is needed, record the rate, date, and source rather than asking a model to guess.
Impressions are a count; add-to-cart and order-conversion rates are ratios. Record each numerator, denominator, and definition. If a denominator is missing, preserve the reported value and mark its definition unconfirmed; do not compare it or infer a cause.
Do not total a file with a missing required column. For a repeated order ID, check for duplicate export, split order, or separate purchases. Preserve both sources and timestamps when values conflict; mark for confirmation.
After correction, retest the original exception and a normal sample; confirm rerunning a duplicate file does not create a second report. An owner approves any send, write-back, or publication.
Compare the example input with the result and acceptance checks.
Fictional practice files. sales.csv: row 2 order_id=1001, amount=12.50, currency=USD; row 3 order_id=1003, amount=9.00, currency=USD. Exception file duplicates.csv: rows 2 and 3 both show order_id=1002, amount=9.00, currency=USD; whether this is a duplicate export is unconfirmed. note.txt row 1: 38,000 impressions; 8.2% add-to-cart rate; 2.1% order conversion rate; denominators not provided; page speed is being checked.
Use only the supplied information. Separate figure checks, campaign summary, exceptions, and missing information. Preserve amounts, currencies, and rates; do not count the duplicate order as confirmed sales or total an unconfirmed amount. State that denominator definitions are missing and comparability cannot be confirmed. Return JSON with string values; write “not provided” for missing text. Do not guess, send, or publish.
{
"sales_file": "sales.csv rows 2-3; USD 12.50 + 9.00 = 21.50",
"duplicate_file": "duplicates.csv rows 2-3; order 1002 repeated; total pending confirmation",
"campaign_note": "note.txt row 1; 38,000 impressions; add-to-cart 8.2%; order conversion 2.1%",
"rate_definitions": "not provided; comparability unconfirmed",
"publication": "draft; owner review required"
}Why it is included: Split input preparation, specialist work, and final synthesis into dependent nodes that consume earlier outputs.
When it does not fit: A single local CSV check does not need a multi-stage orchestration runtime.
Why it is included: Set attempts, backoff, and an exception allowlist for one unreliable node; do not retry business denials.
When it does not fit: Do not blindly retry validation/business rejections or non-idempotent writes.
Why it is included: Pass a client cancellation signal to the Runner so long tools or workflow nodes stop and the event stream closes cleanly.
When it does not fit: A short report that finishes immediately does not need a cancellation state machine.
Why it is included: Emit a structured input request when a workflow lacks a decision or field, then continue after the user responds.
When it does not fit: When the CSV has all required fields and fixed rules, asking the user adds no value.
Why it is included: Pause an agent run before an external change so a person can inspect and approve the exact action.
When it does not fit: A local draft with no external send/publish action needs no approval integration.
Why it is included: Use ordered session events and state deltas to explain agent state transitions instead of guessing from prompts.
When it does not fit: A one-off local run does not need persistent workflow event history.