Self-Learning Invoice Agent

Self-Learning Invoice Agent is an invoice-processing flow built around the DeepSeek Agent node's learning feature. The agent reads a supplier invoice PDF, extracts the vendor, invoice number, invoice date and grand total, and records them to a CSV ledger through an add_invoice_row tool built from ordinary flow nodes. The first invoice from a vendor is worked out from scratch; the agent then writes the whole procedure down as a learned skill in its memory, and every invoice after from that vendor runs on what it learned.

This is AI and RPA working together: the agent decides what to record, and the flow between Tool In and Tool Out records it.

What Self-Learning Invoice Agent can do

• Pick up the next invoice PDF from ~/Invoices/inbox on every run
• Extract vendor, invoice number, invoice date and grand total from the document
• Record exactly one ledger row per invoice via the add_invoice_row tool into ~/Invoices/accounting.csv
• Learn each vendor's layout once: which line holds the real total, which of the dates is the invoice date, how numbers are formatted
• Reuse the learned skill on every later invoice from that vendor, with no re-derivation
• Archive processed invoices to ~/Invoices/processed
• Sync its memory to the workspace as a git repository, so what one robot learns is available to every robot running this flow

Behind the scenes

The flow lists ~/Invoices/inbox, picks the first PDF, and hands it to the DeepSeek Agent node together with a request naming the exact file. The agent reads the document (its persona, in the Context tab, tips it to pdftotext -layout), decides the field values, and calls add_invoice_row. That tool is plain flow: Tool In receives the call's parameters, a Function node shapes them into a one-row table, Append CSV writes the ledger, and Tool Out hands the result back to the agent. The agent's reply leaves on its response port, the invoice is moved to ~/Invoices/processed, and the run ends.

Because the node's Learning property is On, a turn that worked a new layout out ends with the agent recording the procedure via its learn tool. Everything about the agent lives behind a double-click on the node: the learned skill, the memory's commit history, version tags, and a Growth view that replays how the memory grew. A learned skill can be promoted to reviewed, pinned to a digest, and tagged as a version; a robot whose Memory Version names that tag replays exactly the reviewed state and records nothing new.

Setup Guide

1. Create the working folders and the ledger (the ledger needs its header row to exist): mkdir -p ~/Invoices/inbox ~/Invoices/processed echo "vendor,invoice_number,date,total,currency" > ~/Invoices/accounting.csv
2. Configure the API key: put an OpenRouter API key into a Vault item (type: API Key) and select it in the Invoice Agent node's API Key property. The node's default Base URL and model (deepseek/deepseek-v4-pro) already match OpenRouter.
3. Drop an invoice PDF into ~/Invoices/inbox.
4. Run the flow on a robot and watch the turn stream in the Dev Console's Agents tab.
5. Run it again with a second invoice from the same vendor: the skill is offered back to the agent and nothing is re-derived.
6. Inspect the memory: double-click the Invoice Agent node and open the Memory tab to read the skill the agent wrote, its commit history, and the Growth view.

Customization

The agent's persona lives in the flow's AGENT.md asset (visible in the agent editor's Context tab): adjust the fields to extract, the exactness rules, or the learning instruction there. The tool's contract is the Tool In node's description and parameter schema; extend the schema and the Build Row function together to record more columns. Swap Append CSV for a Google Sheets or database node to land rows somewhere other than a local file. In the node's properties, Learning and Memory Sync control whether the agent may learn and where its memory lives; set Learning to Replay with a pinned Memory Version on fleet robots that should run exactly what was reviewed.