One agent, two databases, and a loop that collects cash faster every week.
Remit wakes up every weekday morning, reads the aging report, and decides what every past-due invoice needs: a reminder, an escalation, a dispute flag, or deliberate silence. Every decision is recorded for audit before anyone sees it. Every analyst review becomes a lesson. A weekly refinement pass turns those lessons into better judgment, with no retraining project.
Every box is a real component. Click any one to see what it does.
Data flows in from the ledger, decisions flow out to the AR team, and lessons flow back into memory, so tomorrow's run starts smarter than today's.
The whole system at a glance
Blue lines carry data in, orange lines carry decisions out, and green lines carry learning. The dashed loop is the point: what the memory table holds today shapes what the context loader assembles tomorrow. Click any box for the detail.
One morning, five moments
This is everything the agent does, shown as a story. It plays on its own. Use the controls to pause, change speed, or jump to any step.
The agent wakes itself and gathers the morning
Why it matters: no analyst has to remember to run anything. The agent starts with today's numbers and every lesson it has ever earned.
It reads every past-due invoice and decides what each one needs
"Beacon promised to pay last week and nothing arrived. They have paid late on 4 of their last 6 invoices. This is past reminders."
"Northwind is only 12 days out and has always paid after one friendly nudge. Keep it warm."
"Juniper always pays by day 30, so a reminder now would only annoy a good customer. Rule #12 says leave them alone."
"Ravenna is brand new, with too little history to judge them. I am not confident enough to act."
Why it matters: the right move is not always to send another email. Escalate, remind, deliberately wait, or admit uncertainty, and every call comes with its reasoning written out.
Every decision is written down first
Why it matters: the call, the confidence, and the written reasoning land in a database that can never be edited, so any decision can be audited months later, including the rejected ones.
An analyst reviews the queue in minutes, and nothing sends itself
Emails sent by the agent itself. It drafts, ranks, and explains. A person approves every send, and every edit becomes feedback.
Why it matters: the agent does the reading, checking, and drafting; the human keeps the judgment and the send button.
It compares its calls to what actually happened, and gets smarter
What this week showed
- Analysts rejected early outreach to new customers 3 times: the agent was too eager.
- Every "wait" on Juniper resolved by day 30: rule #12 keeps being right.
- Beacon paid 2 days after the manager's call: the escalation was the right read.
Why it matters: the lessons are written as plain-language rules a human can read, veto, or roll back, and Monday's run starts one rule smarter. No retraining, ever.
The architecture: a harness, three memories, and an ops loop
Everything inside a single run is ephemeral: the context is rebuilt fresh every morning. What persists lives in databases, and what improves the agent ships like software.
Working memory
Assembled fresh each run: the system prompt, today's ledger snapshot, overnight payments, and the learned rules pulled from memory. Nothing stale accumulates between runs.
Reason + tools
Claude judges each invoice, calling tools mid-thought like get_payment_history,
read_email_thread, and draft_reminder, folding results back into its reasoning.
End-of-loop checks
Confidence below 0.70 routes to a human-judgment queue instead of a drafted action. And the hard rule: the agent drafts, it never sends.
Record + deliver
Every decision is appended to the audit record before the digest is assembled. The digest is the only surface a human ever needs to touch.
What happened
Every dated invoice event: reminders sent, promises made, payments received, disputes raised. The raw material the weekly refinement pass replays.
What is durably true
Customer payment behavior, distilled: "Beacon Health pays late 4 of 6 invoices." Retrieved by relevance when an invoice for that customer comes up.
How to act
The dunning playbooks: how to write a first reminder vs. a final notice, when a call beats an email, what tone fits a 30-year account vs. a new one.
Consolidation runs on a cheaper model. After enough new episodes accumulate, a summarizer agent distills episodic history into semantic facts, so the expensive reasoning model reads three crisp sentences about a customer instead of ninety raw ledger events.
Two loops improve two different things. The weekly refinement pass improves the agent's judgment by writing better rules to memory. The ops loop improves the agent's machinery (prompts, tools, retrieval settings), and every change ships behind an eval gate, the same way software ships behind tests.