Move an invoice hold rule out of the model after inconsistent verdicts
At a glance
Mohamad Namroud keeps AI invoice reading but makes missing-delivery holds a software rule. This workflow is listed under invoice-processing, deterministic-rules, evaluation, human-approval.
His agent matched supplier invoices against received items and agreed prices. After repeated inconsistent pay-or-hold decisions, he kept the model for extraction and matching and put the missing-delivery rule in code.
How it went
The same invoice initially received eight correct holds in ten runs. He reports ten correct decisions in ten runs after the change.
Worth knowing
These are repeated trials of one invoice, rather than a varied production benchmark; human approval remains in the workflow.
Save ideas in this browser. Upvotes count across the board.
Anthropic adds Dreaming (research preview: cross-session memory curation; Harvey saw ~6× completion), Outcomes (grader rubric + iterate; +~10 pts on hard tasks), and lead/specialist multiagent orchestration with webhooks—so Managed Agents can self-improve, self-grade, and parallelize without constant human steering.
An Autoresearch Sentinel demo uses a Python controller and Codex to improve a maze solver while protecting the evaluator and recording accepted versions.