Welcome to Rillet Labs
Rillet Labs is the research arm of Rillet, focused on advancing the capabilities of AI agents in finance and accounting. Part of our core mission at Rillet is to be the translation layer from the frontier of AI to the domain of finance.
Our research explores what it takes for AI agents to move beyond simple automation and tackle complex, judgment-heavy accounting work. From understanding accounting intent to investigating transactions and producing verifiable results, we're studying the challenges that stand between today's AI systems and truly autonomous accounting.
Through Rillet Labs, we'll publish our research, experiments, and technical findings, including what works, what fails, and what we learn along the way. This is our first publication.
The cost of a penny
In accounting, two principles are non-negotiable. The first is accuracy: a number is either right or wrong, and "close" counts as wrong. The second is traceability: every conclusion has to lead back to the evidence that supports it. An agent that works on the books has to meet both.
At the end of every month, accounting teams reconcile the ledger balances used for their financial reports. They cross-check supporting records, investigate differences, and determine whether corrections are needed before finalizing the numbers.
For accounts receivable (the money customers owe the business), that means comparing the ledger balance with the invoice register, which lists unpaid invoices by how long they have been outstanding. Any anomalies require an explanation.
In a research evaluation, we asked two agents built on different architectures to investigate the same accounting scenario: a $5,310.50 discrepancy at the end of September. One attributed the discrepancy to a $5,248.01 unapplied payment, and the other identified the cause as a $5,248.00 open credit memo (credit memos reduce how much a customer owes).
The answers are a penny apart, and only one is correct. Examining these two agents put that penny in a different light.
Two kinds of agents
The first design is multi-agent-architecture: a coordinating agent hands work to specialized subagents (here, a reports specialist and an accounts receivable specialist), each working with its own context.
The second is a single-agent architecture with a small standing prompt, which loads accounting guidance as skills when it needs them and keeps the original question in the same context as every finding.
The changing hands
In the multi-agent run, the coordinating agent worked in tandem with two sub-agents, a reports specialist and an accounts receivable specialist. The reports specialist went first, discovering the $5,310.50 difference between two reports and informing the coordinator of the discrepancy amount between the two balances, but did not identify a source. The coordinator then asked the accounts receivable specialist to search for that amount amongst the “receivables data” and “subledger items”.

This is analogous to having a person compare reports that are supposed to be generated from the same data, but aggregate the numbers in different ways. If that person notices an amount in the two reports that should match, but don’t, another person with more expertise in the specific area is asked to look at the original data to explain why.
Each step looks reasonable on its own. The problem sits in between: by the time the question reached the accounts receivable specialist, it had changed from "why do these two reports disagree?" to "which record is close to this amount?". The coordinator had also assumed the cause would be in the subledger before any evidence supported that, and the specialist had no way to challenge it. So it found a payment within $62.49 of the target and treated it as the likely explanation without fully reconciling the difference.
The September increase
The single agent started from the same $5,310.50, then repeated the comparison for earlier months.
| Change in the reported difference | Amount |
|---|---|
| Difference at the end of August | $62.50 |
| Increase during September | $5,248.00 |
| Difference at the end of September | $5,310.50 |
Engineers will recognize this move. When a regression shows up, you find the last point where things worked and look at what changed since. Comparing across dates gave the agent a period (September) and an exact amount ($5,248.00) to investigate. It had no plan at the start: each lookup was chosen from what the previous one returned.
From there, it searched the general ledger journal entries on the AR account for September and found a $5,248.00 credit memo applied to an invoice that had never been billed. An unbilled invoice does not appear in the invoice register, so the memo reduced receivables in the ledger with nothing in the register to reduce. As a result, the gap widened by exactly the amount of the memo.
Why the wrong answer looks right
Both explanations add up. $5,248.01 plus $62.49 is $5,310.50, and so is $5,248.00 plus $62.50. This is where the two principles do their work:
- Accuracy: the $62.50 was already there at the end of August. The $62.49 was never observed anywhere, and only exists to fill the space between the payment and the total.
- Traceability: the credit memo comes with a reason why the ledger moved and the invoice register did not. The payment explanation never showed how an unapplied payment would affect one report and not the other.
Context engineering for financial reasoning
Splitting work across subagents works well for independent tasks, like summarizing separate documents. Accounting isn't like that. It's a relational web of ledgers, invoices, and subreports where a record in one table alters the state of another.
Reading the two runs side by side shaped how we manage context for accounting agents:
Upfront routing locks in bad priors
When an orchestrator splits a task among subagents, it guesses where the root cause lives before inspecting the data. Once scoped to a narrow lane, a subagent's isolated context makes it hard to pivot when evidence points elsewhere.
Because discrepancies span tables, agents need freedom to traverse records dynamically. Keeping execution in one shared context lets each query result inform the next lookup.
Decouple static domain logic from working state
Stuffing every accounting rule, schema, and tool into a massive system prompt degrades attention over long runs, a failure mode called context rot. Leaving out domain rules, however, causes invalid accounting assumptions.
We solved this by separating reusable skills from active state:
- Load accounting skills on demand: We modularize domain logic, like aging bucket calculations or credit memo behavior, into discrete skills. Through progressive disclosure, the agent loads skills only when needed, keeping the base prompt lean.
- Keep discovered state pinned: Case-specific findings, like isolating the $62.50 August carryover to establish September's $5,248.00 delta, stay pinned in the working context so intermediate math stays intact.
Context compaction must preserve causal lineage
Long investigations hit context limits, forcing the harness to compact earlier steps. Standard summarization optimizes for token reduction, stripping out state needed to verify math downstream.
If compaction retains only "found a $5,248.00 credit memo," the agent loses the causal chain:
- How that memo explains September's delta once you factor out the $62.50 August balance.
- Why it reduced the general ledger without changing the invoice register, given the invoice's unbilled state.
Context compression is a domain-dependent task: the system must identify which details and decisions later work will need. In accounting, a conclusion without an audit trail isn't usable. After a handoff or a compaction, the agent must still be able to support its math, point to the records, and say what remains unresolved.
$0.01 matters
$5,248.01 or $5,248.00. At a shop counter, a penny goes in the tray and nobody thinks about it. At Rillet, we care about that $0.01.
Accounting evidence can span customer records, ledger entries, earlier periods and reporting rules. We must be precise about the entity, period, currency and records it covers. An almost correct amount can easily lead us astray.
That understanding shapes our approach to the harness. For us, teaching the agent accounting turned out to be the smaller part of the work. The larger part is deciding what the agent is allowed to forget, so that its conclusions stay accurate and traceable through a handoff, a compaction, or a long run.
That is what we are building at Rillet: agents whose conclusions accountants can verify and defend. The take-a-penny tray can stay at the shop counter.
