Long financial tables lose rows, legal files scatter related clauses across thousands of pages, and diagrams hide text in fine print. Extend's Operator-1 is an autonomous agent built to extract data from documents like these and check its own results.
Extend is a New York-based Document AI company. It provides models, APIs and infrastructure to parse, extract, split and edit documents, run evaluations and orchestrate workflows.
Extend says extraction systems that rely on chunking documents and a single model pass struggle with this material. Rows get skipped or duplicated, and related information is missed.
Operator-1 takes a different approach. It decides which steps each extraction task requires, then works through them in a dedicated workspace with document processing tools and persistent memory.
The agent can write and run code to extract table rows, transform data and validate calculations. One example is comparing the sum of extracted line items with the total reported in a financial statement.
It can search across documents using command-line and custom search tools. That lets it find and connect information spread across thousands of pages.
A crop and zoom tool inspects specific regions of document images. The agent can examine page breaks or link the details in a diagram with the surrounding text.
Each result is checked against the task's requirements. When a check fails, Operator-1 triggers another attempt.
Successful code and strategies are kept in tenant-specific memory and runbooks. They are reused on later documents of the same type, which Extend says reduces repeated work and may lower runtime costs.
Extend reports that Operator-1 scored 99.3 per cent on LongArray-Extract, its own open-source benchmark for extracting long arrays from real-world documents.
It scored 99.88 per cent on LongExtractionBench, a micro1 benchmark commissioned by Reducto that tests filling JSON schemas from long documents. On LlamaIndex's ExtractBench, covering structured extraction from business documents, it scored 96.57 per cent.
Each benchmark uses a different dataset and evaluation method, and two of the three were published by third parties. Extend says the results set a new record on all three and show the benchmarks have been saturated.
Kushal Byatnal introduced Operator-1 for Extend on X. He described it as built for long tables with thousands of rows, files with context scattered across hundreds of pages, and dense diagrams.
Operator-1 is available now in the Extend dashboard and API, and is included on all plans, including PAYG, Scale and Enterprise. It is charged at a fixed five credits per page.
New users can test it with 10,000 free credits.