Search COMPUTER USE

Search the evidence-backed index by name, vendor, or use case.

COMPUTER USE / WORKFLOW

Reconcile CSV rows against a web portal

Match by explicit keys, retain observed values, and review discrepancies without changing the portal.

Before the first run

  • A validated input CSV with stable keys
  • An authorized read-only portal scope
  • Defined comparison rules including units and dates
ACTION BOUNDARY

Define what the agent can do.

Read and compare only. Correcting a portal value or updating the source file’s authoritative record is a separate task.

The implementation sequence

  1. 01

    Validate the file

    Check headers, required keys, duplicates, and type parsing before opening a browser. Preserve the raw input alongside normalized values.

  2. 02

    Match one record at a time

    Use the approved identifier set to locate the portal record. Return an ambiguous state when more than one plausible record appears.

  3. 03

    Capture comparable values

    Record the observed field, unit, currency, date, and source reference. Do not compare a monthly total with a year-to-date total simply because their labels look similar.

  4. 04

    Apply explicit rules

    Run deterministic equality or tolerance checks outside the agent. Preserve both values and the rule that produced each discrepancy.

  5. 05

    Review the exception report

    Separate matches, mismatches, missing records, and inaccessible records. The report should let a reviewer inspect one discrepancy without repeating all lookups.

Stop for human approval

Authorize the record scope and export destination; a responsible reviewer decides what action to take on discrepancies.

Approval design principles →

Implementation options

Failure & recovery

  • Locale-dependent CSV parsing
  • Ambiguous record matches
  • Incompatible units or periods
  • Portal stale data

What drives cost

  • Rows and lookup latency
  • Pagination and search complexity
  • Mismatch review
Estimate your stack cost

Related use cases

RESEARCH NOTES

Sources & verification

Reviewed Sep 27, 2026. Architecture recommendations are editorial analysis; linked vendor documentation supports the underlying capability and safety facts.

  1. playwright.dev/docs/locators
  2. playwright.dev/docs/auth
  3. developers.openai.com/api/docs/guides/tools-computer-use
How we verify evidence →