Extraction agents read documents, invoices, receipts and contracts, and turn them into fields: amounts, dates, parties, terms. Then they decide where each field goes and write it to the right system.
Reading a messy document is real model work, and it stays with your agent's model. What comes after it is the same every time: where the fields go, what to check, what to write. That's the part that runs compiled.
What you get
- Your model reads the document, as it does today.
- The steps after it run compiled, the same way every time.
- Anything that doesn't fit goes to your agent, unchanged.
This sits close to what AgentCompile does today, and it's where we're looking next.
Two kinds of work in one agent
An extraction agent does two different things. Reading is the hard part: a scanned invoice, a receipt photographed at an angle, a contract with clauses in odd places. That's real model work and it stays with your model. Then comes the routine part: check the vendor exists, match the purchase order, file the fields, flag anything that doesn't add up. That part is the same every time.
Without AgentCompile
- The model reads the document, then reasons through every step after it.
- Each invoice is filed by a fresh round of reasoning.
- The routine part costs as much thought as the reading.
With AgentCompile
- The model reads the document, exactly as today.
- The steps after it run compiled, the same way every time.
- Anything unusual (a new vendor, a mismatch) goes to your agent.
An invoice, walked through
- A new invoice arrives. Your model reads it into fields, as it does today.
- Look up the vendor, match the purchase order, check the amounts: this part repeats, so it can run compiled.
- Writing to your accounting system is a change, and it follows your approval rules.
- An invoice that doesn't match goes to your agent, with everything found so far.
Where we are
This is close to what AgentCompile does today, and it's where we're looking next. We haven't published results for extraction agents yet.