Applied Patterns

The standard agent shape

Every agent in this chapter follows the same spec template. Use it in your interview answers — it signals platform thinking.

1 — Reconciliation agent

The workhorse. Reconciles a balance-sheet account between NetSuite (book) and a source system (bank, custody, exchange, intercompany).

The core matching loop

Matching is deterministic, not LLM-driven. The LLM's job is around the matching:

1. Pull GL detail (all transactions in account for period).
2. Pull source detail (bank statement, custody ledger, exchange trade log).
3. Deterministic match: exact match on amount + date ± tolerance, then on reference id, then fuzzy on description.
4. LLM step on the unmatched residual: "given these N unmatched GL entries and M unmatched source items, propose plausible pairings, missing entries, or timing differences."
5. LLM proposes reconciling items with evidence citations.
6. Human reviews proposals in BlackLine; approved items become JEs.

2 — Treasury monitor

A daily (or intraday) agent that watches cash and crypto positions, flags thresholds, drafts the treasurer's morning brief.

Why this is a great first ship

Tier 1, high visibility, treasurer is the immediate user. If you ship one production agent in your first 90 days, this is the safest candidate. Show the treasurer their morning brief was 5 minutes faster and contained one anomaly callout they'd have missed — you've earned trust for the next, riskier build.

3 — Audit-prep agent

External audit and quarterly internal audit eat finance time. This agent prepares evidence packages on request.

The killer feature: the agent's own runs are themselves in the audit log. When an auditor asks "how did you produce this package?" you can replay the agent's actions.

4 — Controls-validation agent

SOX requires evidence that controls operated effectively. This agent samples transactions, evaluates them against the control narrative, and produces the testing workpaper.

The reproducibility (random seed) detail is high-signal — it tells auditors they can rerun your sample and get the same selection.

5 — Close-checklist orchestrator

The planner-style agent that runs the close. Not a doer; a coordinator. It tracks the 80-task close list, nudges owners, summarizes blockers, and dispatches the recon agents.

This is the cleanest first instance of the platform pattern: one agent coordinating other agents, each with its own Process Owner.

6 — Variance commentary agent

At close, every account with material movement period-over-period gets a commentary ("Revenue is up $4.2M vs. prior month driven by..."). This is grinding work; LLMs are very good at it.

Add an anti-fabrication rule in the system prompt: "Every numeric claim must cite a transaction id or a balance fetched via tool. Do not invent reasons. If you cannot identify a driver, say so."

Composing agents — the platform view

These agents are not islands. A real close looks like:

  T-3: Close orchestrator starts the calendar.
  T-2: Pre-close checklist nudges. Treasury monitor flags cash positions.
  T-1: Recon agents fan out across 40 accounts.
       Variance commentary drafts as recons complete.
  T+0: Controller reviews proposed JEs. Approves the recurring accruals.
       Approves variance commentary.
  T+1: Audit-prep agent assembles the close package for external auditor sample.
       Controls-validation agent samples March JEs > $500K and produces workpaper.
  T+2: Close signed off.
  

The audit log spans all of this. Every agent's run is queryable: "show me all proposed JEs that were rejected by the controller in the last 6 closes" becomes a one-line query — and that's exactly the kind of question an auditor will ask.

Things to prototype before interview day

If you have 1-2 evenings, build one of these. It will dramatically sharpen your answers.

• Tiny recon agent: two CSVs (a "GL" and a "bank statement" with deliberate breaks), Claude API, three tools (load_gl, load_bank, propose_match). Have it output a structured ReconciliationResult.
• Treasury monitor: a mock balances JSON, a Claude call that produces a morning brief, an n8n flow that wraps it in a Slack post.
• MCP server skeleton: a stdio MCP server that exposes two tools (fetch_balance, list_accounts) against a SQLite fixture.

Each is 60-120 minutes. The shared payoff: you can say in your interview "I built this last week, here's what I learned" — and that beats any amount of theoretical answer.

✓ Check yourself

Answer from memory before opening each one.

1. A reconciliation has 4,000 ledger lines and 3,950 bank lines. Where in the process does the model take part, and why not earlier?

   Show answer

   The model works only on the residual after deterministic matching on amount, date, reference and a fuzzy description pass. Left to match everything, it will confidently pair items that do not belong together and you cannot prove otherwise. Deterministic matching is exact and repeatable, so the model's judgement is spent where rules have run out. See the reconciliation agent.

2. You can ship only one of the six agents in your first 90 days. Which, and what is your reasoning?

   Show answer

   The treasury monitor is the safest candidate: it is Tier 1, has an immediate named user in the Treasurer, and never moves money. A brief that is faster and catches one missed anomaly earns the trust that riskier builds need. If the Controller is the sponsor and close pain is bigger, the close status orchestrator is the alternative, also Tier 1. See the treasury monitor.

3. Why does the controls-validation agent log the random seed of its sample?

   Show answer

   Sampling must be reproducible: with the same seed and the same universe, an auditor or the SOX lead can rerun it and get the same selection. Without the seed, the workpaper cannot be re-performed and the sample could be dismissed as cherry-picked. See controls validation.

4. An auditor writes: send everything unusual from the spring. What should the audit-prep agent do?

   Show answer

   Stop and ask clarifying questions: which dates, which entities, which accounts, what threshold counts as unusual. A guess produces a package that looks complete but answers a different question. Once scoped, the manifest records the request text and the exact queries run. See the audit-prep agent.

5. The variance agent's prompt says every number must cite a transaction or a fetched balance. How do you make sure that rule holds, not just hope?

   Show answer

   Check it in code: parse the citations from the draft and confirm each one exists in the tool results, then reject or flag any unmatched claim. A prompt rule is a request, while a check on the output is a control. The recon owner still reviews before close. See the variance commentary agent.

6. The close orchestrator is Tier 1, and it spawns recon agents that propose entries. Does the whole workflow count as Tier 1?

   Show answer

   No. Each sub-agent keeps its own tier, Process Owner and gates, so the controls follow the agent that can write, not the one that coordinates. Reading the orchestrator's tier as the workflow's tier would hide the riskiest step. See the close orchestrator and composition.

Takeaway

• Specify every agent the same way: goal, Process Owner, tier, inputs, tools, output, human gate, audit fields and bail conditions.
• Reconciliation is deterministic first and model second: rules match what they can, and the model works the residual and proposes items with evidence.
• The treasury monitor is the safest first ship, because it is Tier 1 with an engaged user and never moves money.
• Audit-prep and controls-validation agents are read-only, bail out on ambiguity, and record the original request and the sampling seed so work can be reproduced.
• An orchestrator may be Tier 1, but each sub-agent keeps its own tier, owner and gates.
• Anti-fabrication needs a check on the output, not only a line in the prompt: every number cites a fetched source.
• Agent runs share one audit log, so a question across a whole close is a single query.

Specifying the agents is half the work. The other half is proving they are right, which is the subject of evaluation. → Evaluation & Quality

Further reading

• Building effective agents — Anthropic, 2024. Orchestrator and worker, routing and evaluator patterns behind these specs.
• How we built our multi-agent research system — Anthropic, 2025. Coordinating sub-agents and what it costs in production.
• MCP tools — Model Context Protocol project, 2025. Defining the tool surface each agent calls.
• AS 1215: Audit Documentation — PCAOB. What audit evidence and workpapers must contain, which sets the bar for an evidence package.
• AS 2315: Audit Sampling — PCAOB. How samples are selected and evaluated, which a controls-testing agent must respect.
• Tool use with Claude — Anthropic docs, accessed 2026. The call and result loop each agent runs.