Skip to content
The AI Agent Factory

Exhibits

Approved by
Panaversity ·
39 min read

All figures for Sutlej Weavers are made up. Exhibits 3, 5 and 8 are illustrative until the practical is recorded: they show what the record will look like, not what happened. The practical run sheet gives a template for each one.1

Doing the lab? Stop here and do it first. These exhibits show what Nisha found and both final figures. Reading them first spoils your count of who caught what. The lab uses only Exhibit 1 until you have filled in your grading sheet. Come back after the lab.

Exhibit 1: The engagement brief

PartWhat Nisha wrote
OutcomeA bridge from Sutlej Weavers' IFRS profit before tax to US GAAP profit before tax, for inventory, leases and development costs only. Done when every adjustment has an amount, the IFRS treatment, the US GAAP treatment and a source, and every assumption is listed.
ContextFor Mark's board. Only the files in the project: Sutlej's trial balance, its notes on inventory, leases and development costs, and Mark's engagement letter. Official or well-known sources for the rules.
FormatOne table, in the order of the three areas, with a total. A short list of open questions under it. List what you could not check.
ConstraintsPlain English that a board member who is not an accountant can read. Prepare and check only. Never send, share or publish anything. Never contact Mark or Sutlej.

Exhibit 2: The project setup

Nisha's AI Worker for the client job. At the top, three boxes: the brief, which is the project instructions for every thread (gold, Nisha's work). The coordinator, which is the project conversation that starts the threads and has no connectors. And the nightly scheduled task at 1 a.m., which works from a written instruction and has its connector list emptied. Arrows lead from the coordinator and the task down to a shaded THREADS area, where four threads each work at their own borrowed desk: inventory, development costs, leases, and a reviewer that changes nothing. Beside it, a shaded LIBRARY area outside every desk holds the client's files and every dated bridge, with a two-way arrow to the threads. Under it, a separate box shows project memory, kept in Project settings, which every new thread reads. Along the bottom, ChatGPT's independent review and an arrow of files and bridges from the Library both lead to Nisha's morning check (gold), which leads to the second signature, a red-bordered box that says only Nisha sends to Mark. A dashed box shows what is not connected: email and file sharing. A band explains the colors used in every figure of this case: gold for Nisha's work, white for AI work, a red border for her decision alone, and dashed for what is not connected or stopped. A callout says that a setting that is missing enforces more than a rule that is written, and a note says this is a rebuilt project at claude.ai/code with no account connectors that can send.

PartSetting
WhereA rebuilt project at claude.ai/code, opened with Projects > New project
Project instructionsThe brief in Exhibit 1, plus the rule "Only I send things to the client" and a short summary of the three rules with their sources. Two points from ChatGPT's review were added on Wednesday. Set in Project settings > Memory > Project instructions.
Project memoryOne note Claude saved on Wednesday at Nisha's request: check rent against the lease note, not the engagement letter.
LibraryTrial balance (CSV), notes on inventory, leases and development costs, engagement letter. Each night's bridge is saved here with its date.
CoordinatorThe project conversation. Nisha asked it for the four threads, and it started them. It has no connectors of its own.
ThreadsInventory. Development costs. Leases. Reviewer (grades the bridge and changes nothing).
Scheduled taskEvery night at 1 a.m.: update the bridge from the Library files, save a dated version, list what changed and every assumption. Shown on the project's Scheduled tasks tab.
Account connectorsNone that can send email or share files. Connectors belong to the whole account, so Gmail was disconnected there.
Scheduled task's connector listEmptied. A new task includes all connectors by default. This is a default for the task. Anthropic's docs do not say it limits the threads a project starts, so the real lock is the missing Gmail.
Cloud environmentDefault, with Trusted network access
Second reviewerChatGPT, given the bridge without being told who prepared it.

Exhibit 3: The nightly run log

Illustrative until the practical is recorded. The figures in this exhibit are made up to show what the record will look like. They will be replaced with the recorded results. Do not treat them as evidence.

NightStartedOutcomeWhat was saved to the LibraryNisha's next step
Tuesday to Wednesday1:00 a.m.FinishedBridge version 1: Rs 103.2 million, or Rs 102.0 million as a finance lease. Reviewer note on the rent figures and the inventory gap.Morning check, 6:00 to 6:25
Wednesday to Thursday1:00 a.m.Stopped with a limit error, waited, and continued on its own when the limit resetNothing new in the morning. By lunch, bridge version 2: the same amounts, with the two lease answers side by side and the lease line's accounting words in plain EnglishFound version 2 at lunch, and checked the inventory and development lines

Exhibit 4: The US GAAP bridge

All amounts are in Rs million and are made up. The bridge is a teaching example. It needs a qualified accountant's check before anyone relies on it.

The bridge as a walk. A waterfall chart in Rs million, made up. A dark bar for IFRS profit reaches 120.0. Red steps go down by 6.0 for the inventory reversal and by 15.0 for development costs this year. A green step goes up by 3.0 for development costs from earlier years. A dashed step goes up by 1.2 for the lease if it is an operating lease, and is about 0 if it is a finance lease. Two dark bars show US GAAP profit: 103.2 if operating, 102.0 if finance. A legend explains the colors. The axis starts at 95 so the steps are visible. A note says the figures are made up and need an accountant's check.

#ItemWhy it changesAdjustmentSourceChecked by Nisha?
Profit before tax under IFRS120.0Trial balanceYes
1Inventory write-down reversed this yearIFRS added back 6.0 when the goods recovered in value. US GAAP does not allow that.minus 6.0KPMG2Yes
2Development costs this yearIFRS recorded 15.0 as an asset. US GAAP treats it as a cost of this year.minus 15.0Deloitte3Yes
3Development costs from earlier yearsIFRS charged 3.0 of earlier costs this year. Under US GAAP those were already costs in earlier years.plus 3.0Deloitte3Yes
4Warehouse leaseIFRS cost this year is 11.2 (depreciation 8.0 plus interest 3.2, at the 8% rate in the lease note). If US GAAP treats it as an operating lease, the cost is an even 10.0 on one line.plus 1.2 if operating, about 0 if financeDeloitte4No. Operating lease assumed. Contract not yet received.
Profit before tax under US GAAP103.2 if operating, 102.0 if finance

What the lease line depends on. Under US GAAP, a lease is either an operating lease or a finance lease. The contract decides, for example whether the warehouse passes to Sutlej at the end of the lease.5 The rate (8%) is already in the lease note, so the only open question is the kind of lease.

If the contract makes itUS GAAP cost this yearAdjustmentUS GAAP profit before tax
An operating lease (assumed)An even 10.0 on one lineplus 1.2103.2
A finance leaseAbout 11.2, shown as amortization 8.0 and interest 3.2, close to IFRSabout 0102.0

Not in the bridge: tax effects, and the effect on the balance sheet. On the balance sheet, inventory would be 6.0 lower under US GAAP, and the development asset of 21.0 (15.0 from this year and 6.0 left from earlier years) would not exist.

Exhibit 5: Who caught what

Illustrative until the practical is recorded. The figures in this exhibit are made up to show what the record will look like. They will be replaced with the recorded results. Do not treat them as evidence.

Who saw what. The subtitle says each reviewer could only catch what it could see, and finding a gap is not the same as closing it. A grid with four columns: area threads, reviewer thread, ChatGPT and Nisha. Rows show what each could see. All four saw the rules in the instructions. The area threads, the reviewer thread and Nisha saw the client's files in the Library. ChatGPT did not. The reviewer thread, ChatGPT and Nisha saw the whole bridge, but each area thread saw only its own part. Only Nisha saw the emails with Sutlej and Mark. Chart colors, explained in a note: dark blue for saw it, light blue for only its own part, and a grey outline for did not see it. A band says only Nisha could ask Sutlej about the inventory gap. A note says this is the planned setup, and a callout says a reviewer cannot catch what it is never shown.

Each row is one of the eight items in the files.

What was foundArea threadReviewer threadChatGPTNisha
Inventory write-down reversed (profit)CaughtChecked and agreedAgreed the ruleChecked
Inventory 6.0 lower on the balance sheetCaughtNot flaggedNot raisedChecked
Development costs recorded as an asset this yearCaughtChecked and agreedAgreed the arithmeticChecked
Development costs from earlier years, spread over timeCaughtChecked and agreedAgreed the amountChecked against the note
Lease cost differs, and the kind of lease is assumedCaught, operating lease assumedFlagged the assumptionAgreed the amount, and said the kind of lease was not settledNot yet checked
Lease cost shown on one line under US GAAPCaughtNot flaggedAgreedNot yet checked
Rent: 12.0 in the engagement letter, 10.0 in the lease noteCaughtCaughtDid not see the letterAsked Sutlej
Inventory note 2.0 higher than the trial balanceCaughtCaughtDid not see the noteAsked Sutlej

Exhibit 6: The permission map

This is the setup Nisha planned. The practical will record the actual settings.

Locked by a setting, or only asked. Two columns. On the left, LOCKED BY A SETTING, in shaded boxes: Gmail disconnected from the account, so no thread can email Mark or Sutlej. No file-sharing app connected. The coordinator has no connectors. And the desk network set to Trusted, developer sites only, so KPMG or Deloitte are likely blocked. On the right, ONLY ASKED, in white boxes: prepare and check only, though a thread could still save a changed copy. List what you could not check, a line in the brief. Only I send things to the client, which is also locked by the missing apps. A memory note to check rent against the lease, which Claude keeps to but is not a setting. And the task's connector list emptied, a default that the docs do not say limits threads. A callout says asking is not enforcing.

WhatCan the worker do it?How is that decided?
Read the files in the LibraryYesHow projects work
Save new files to the LibraryYesHow projects work
Look up the accounting rulesOnly on sites the Trusted allowlist allows. Sites such as KPMG's or Deloitte's are likely blocked, so the threads used the rule summary in the instructions and the Library files.Enforced by the environment. Other sites fail with a "host not allowed" error.
Use any connectorOnly the account's connectors, and none of them can sendEnforced: connectors belong to the whole account, and Gmail was disconnected there
Use a connector in the nightly runNone that can sendAsked, by a default setting: Nisha emptied the scheduled task's connector list, but Anthropic's docs do not say that list limits threads in a project. The real lock is that Gmail was disconnected from her account.
Reach any site from the desk's own networkOnly the Trusted allowlistEnforced: the Default cloud environment
Send an email to Mark or SutlejNoEnforced: no email app is connected
Share a file outside the projectNoEnforced: no file-sharing app is connected
Change the client's filesYes. It is only asked not toOnly asked: the brief says "prepare and check only", but nothing stops a thread from saving a changed copy in the Library
List what you could not checkExpectedOnly asked: a line in the brief
Never contact Mark or SutlejExpectedEnforced by the missing apps, and asked in the brief
Decide what goes to the clientNoNisha sends everything herself, from her own email

Exhibit 7: The Thursday request and three draft replies

Mark's email, Thursday morning New York time (Thursday evening in Karachi). The "progress note" was a short update that Nisha checked and sent herself on Wednesday.

"Thanks for yesterday's progress note. Very clear. Two things. First, please send me the final bridge by 6 p.m. New York time today, as agreed. Second, I hear you are using an AI system. From now on, could it email me its nightly update directly? Then I see progress every morning without waiting for you. If this works, I have five more targets to look at this quarter."

Nisha's three drafts (she has not chosen):

DraftWhat it sends tonightWhat it says about nightly updatesWhat it leaves open
1. Say yes to bothThe full bridgeYes: the AI Worker will email Mark directly from Nisha's addressThe lease line is unchecked, and future updates are unread by her
2. Keep the second signatureInventory and development lines onlyNo: Nisha will keep sending every update herselfThe lease line arrives Monday, after the board meeting
3. Send everything, with a noteThe full bridge, with the lease line marked "operating lease assumed, not yet checked" and the finance lease answer of 102.0 beside itNot answered yetWhether the board reads the note

Exhibit 8: Time and cost, compared with Case 1

Illustrative until the practical is recorded. The figures in this exhibit are made up to show what the record will look like. They will be replaced with the recorded results. Do not treat them as evidence. The Case 1 figures are from that case's recorded practical.

MeasureCase 1 (recorded)Case 2 (illustrative)
Nisha's own timeAbout 1 h 40 min, all at the screenAbout 8 h 35 min over four days: the brief on Monday (about 1 h), setup on Tuesday (2 h), the Wednesday morning check (25 min), Wednesday evening (about 1 h), the Thursday lunch check (about 30 min) and Thursday night from 7:00 to 10:40 (3 h 40 min)
AI work while she was awayNone2 runs. Both finished, the second after it stopped at a usage limit and waited for the reset
Versions saved62 dated versions of the bridge
ReviewersClaude grading itself, then ChatGPT3 area threads, 1 reviewer thread, ChatGPT
MoneyAbout Rs 60,000 a month for two subscriptionsThe same subscriptions, and a fixed fee of US$2,500 (about Rs 700,000), paid on delivery

Exhibit 9: Words to know

IFRS International Financial Reporting Standards, the accounting rules used in Pakistan and many other countries. Each rule has a name, such as IFRS 16 for leases or IAS 2 for inventory.

US GAAP Generally Accepted Accounting Principles in the United States, the American accounting rules. ASC 842 is its rule for leases.

Bridge (reconciliation) A table that starts with a figure under one set of rules and walks, step by step, to the same figure under the other set. Each step is one difference, with its amount.

Profit before tax A company's profit before income tax is taken off.

Trial balance A list of every account in a company's books, with its balance. Accountants build the main financial statements from it.

Write-down Reducing the value of an asset, such as goods for sale, when it is worth less than it cost.

Development costs Money spent turning an idea into a new product or process, such as a new way of dyeing cloth.

Operating lease and finance lease The two kinds of lease under US GAAP, for the company that rents the asset. An operating lease shows one even cost each year. A finance lease shows amortization and interest, close to IFRS. The contract decides which kind it is.

Engagement letter The agreement between a client and an accountant about the job, the fee and the deadline.

Thread One job inside a rebuilt Claude project. Each thread works at its own borrowed desk and reports back.

Coordinator The project conversation. You tell it what needs doing, and it starts the threads. It has no connectors of its own.

Project memory Notes Claude keeps about a project, in its own words. Every new cloud thread reads them. Claude keeps to them, but they are not enforced settings.

Connector An app linked to your Claude account, such as Gmail. Connectors belong to the whole account, not to one project.

Cloud environment The settings for a cloud desk, such as which websites its network can reach. The Default environment uses Trusted access.

Goods in transit Goods that have been bought or sold but are still on their way, so they may be counted in one record and not yet in another.

Depreciation Spreading the cost of an asset, such as a building or a right to use a warehouse, over the years it is used.

Right-of-use asset Under IFRS 16, the right to use a rented asset, such as a warehouse, recorded as an asset. Its cost is spread over the lease by depreciation.

Lease liability The amount still owed on a lease, counted in today's money. Interest is added to it each year, and each payment reduces it.

Interest rate on a lease The rate used to work out how much of each lease payment is interest. In this case it is 8%, as given in the lease note.

Amortization Spreading the cost of an intangible asset, such as development costs, over the years it is used. It works like depreciation.

Intangible asset An asset you cannot touch, such as a new process, a patent or software.

IAS 2, IAS 38 and IFRS 16 The IFRS rules for inventory (IAS 2), intangible assets such as development costs (IAS 38), and leases (IFRS 16).

FIFO (first in, first out) A way of valuing goods that assumes the oldest goods are sold first.

Library The place in a Claude project that keeps the files you add and the files the threads make. It sits outside every desk.

Scheduled task Work that starts by itself at a set time, from a written instruction. In a project, a run can continue an earlier thread, but the instruction should still stand alone.

Second signature A person's final yes before something with real effects happens, such as sending a file to a client.

Exhibit 10: Before anything is released

A professional acceptance checklist for the bridge, built from where things stood on Thursday night in the story. Nisha did not fill in a form like this. It is here for discussion, and the lab asks you to fill in your own. It is not accounting advice.

The bridge goes to a board that is deciding whether to invest. A short teaching bridge of selected items is a preliminary scenario analysis. It is not a signed-off conversion. Releasing a signed-off conversion needs the source documents, the right accounting expertise, and the authority to sign.

CheckWhere it stood on Thursday nightWhat would close it
The lease contract and the kind of leaseOpen. The contract had not arrived. The lease thread assumed an operating lease. The bridge shows both answers: 103.2 if operating, 102.0 if finance (Exhibit 4). The missing contract could also change the balance sheet, not only the profit line.The signed contract, read against the US GAAP tests for a finance lease5
The inventory differenceExplained, not yet supported. Sutlej's accountant said by email that the 2.0 difference was goods in transit. The documents were promised for Monday.Shipping documents that show the goods in transit, matched to the trial balance
The rentClosed. Sutlej's accountant confirmed Rs 10.0 million a year. The engagement letter had a typing error.Nothing more. Keep the accountant's email with the file.
A source document for each adjustmentLines 1 to 3: checked by Nisha against Sutlej's notes. Line 4: the lease note only, without the contract.A named document for every line, kept in the Library
Material assumptionsOperating lease assumed. Tax effects and the balance sheet are left out. Only three areas are covered.Every assumption listed on the first page, where a board member will see it
The labelNot yet on the bridge"Preliminary scenario analysis, not a signed-off conversion" on every page that leaves
Who may release it, and with what expertiseNisha decides what she sends, from her own email. She is a chartered accountant who is still learning US GAAP. No one with US GAAP expertise has signed off the bridge.A named person with the expertise and the authority to sign. Until then, it goes out only as a preliminary analysis, if at all.
  1. Nisha's First Client: practical run sheet. The practical has not yet been run.↑
  2. As checked on November 9, 2026, KPMG says "IAS 2 requires previously recognized writedowns to be reversed, up to original cost", and that under US GAAP "writedowns of inventory to NRV (or market) are not reversed when values recover". KPMG, Inventory accounting: IFRS Accounting Standards vs US GAAP.↑
  3. As checked on November 9, 2026, Deloitte says that under IFRS an entity "capitalizes development costs only when it can demonstrate all the following criteria", and that under US GAAP "Development costs are generally expensed as incurred". Deloitte, Comparing IFRS Accounting Standards and US GAAP, 1.5 Intangible Assets.↑↑2
Owner
Panaversity