Skip to content
The AI Agent Factory

Exhibits

Approved by
Panaversity ·
34 min read

These exhibits hold the details behind the story. Read them when a story page points here.

About this case, in full

  • The story is fiction. The firm, the client, the bank, the people, the figures and the times are made up.
  • The product facts were checked. What each part of Claude can do comes from Anthropic's help pages, checked in November 2026 for this case and for the earlier sections it builds on. Each fact has a note with its source.
  • The joining was not replayed. Nobody ran this exact setup on a Team plan for this case. The run times (25 minutes, 6 minutes) and the line counts are illustrative. The practical run sheet says how to test it.1
  • Eight things the story assumes. A chat in a project can read a connected drive, and save a file there. It runs the matching as a program on the cloud computer.2 A run updates existing artifacts when it is given their links.3 Three people on a Team plan can share one project, one skill and the four pieces. The worker in Slack can run one more routine, as it does in Case 3, and read the latest run note in the drive.4 A sheet can hold the month's 6,800 lines. The lab's free path works as written. The run sheet tests each one.
  • The matching rule is simplified. Real matching also uses reference numbers, payment types and names, and treats every unclear pair as a question for a person.
  • It is not accounting advice. A real bank reconciliation, and a real promise to a lender, have more detail than one figure. A real bank would usually date a returned check in October, the day it came back. The story dates it September 30 to keep the sums simple. And a real promise to a lender is usually tested in a more detailed way than one adjusted bank balance. The $195,000.00 rule here is made up for the case.

Exhibit 1: Where the earlier sections appear in this case

In this caseYou learned it in
Handing over a whole job, and checking what comes backThe New Way of Working
A folder on a computer, read through the desktop appSet up the apps
The temporary computer in the cloudWhere the work happens and Why the borrowed desk gets cleared
Connected apps and skillsClaude 2.0: conversations and artifacts
Scheduled tasksWork that runs while you sleep
Permissions, not wordsStaying safe
Projects (the classic kind)Docs
The sheet, the doc, the deck and the designBuild Something
A worker in Slack with its own nameCase 3

Exhibit 2: The setup, part by part

PartWhat it holdsWhere it lastsPlan
The project, "Lakeshore month-end"The client's rules, three files from Dana, the links of the four pieces. It uses the shared drive, the only connected app Nisha turned on for the firm.5In the firm's Claude accountEvery plan. Free allows five projects.6
Sana's skillThe memo method: READY, NOT READY, unrecorded itemsIn Claude's settingsEvery plan, with code execution turned on7
The shared drive, connectedThe bank's lines, the ledger, the matched list, the run notesIn the driveA connected app
Hamza's folderDana's three files, before they were copied to the project, and the bank's statement for Hamza's own testsOn his computerNeeds a paid plan and the desktop app
The borrowed deskNothing. It runs the matching program and is cleared.2Nowhere
The four artifactsThe sheet, the memo, the slides, the coverIn the firm's Claude accountDocs, slides and designs on every plan. Sheets need a paid plan.8
The routine (a scheduled task)One instruction, and a timeWith the worker in Slack, in the firm's channel4Claude in Slack is on Team and Enterprise plans.9 In Claude itself, scheduled tasks need a paid plan.10

Exhibit 3: The project's instructions

Lakeshore month-end. You prepare the month-end pack for Lakeshore Supply, Cleveland. The pack is for Lakeshore's bank.

Files. The month's bank lines and ledger are in the shared drive, in Lakeshore/ and then the month, for example 2027-09. Use the newest bank file. If it was downloaded more than a day ago, stop and say so. The bank's letter and the layout the bank wants are in this project's files.

Method. Match each bank line to a ledger line with the same amount and a date no more than five days apart. Use the name or check number too, where there is one. If a bank line could match more than one ledger line, pair it with one, and list the pair under Needs a look. List every line that has no match. Work out reconciled cash: the bank's closing balance, plus deposits the bank has not shown yet, minus checks the bank has not paid yet. Save the matched list to the month's folder.

The pack. Update these four pieces, and do not make new ones: the sheet [link], the memo [link], the slides [link], the cover [link]. Write the memo with the month-end memo skill. Slide 2 shows reconciled cash and the bank's minimum, side by side. The bank's minimum is $195,000.00.

Run note. After every run, save a note to the month's folder: the files you read, the number of lines matched, the lines left over, reconciled cash, and the pieces you updated. If this is not the first run for the month, say what changed.

Never. Never send anything to anyone. Never change a file in the drive that you did not make. If the bank and the books do not agree, say so in the first line. Nisha or Sana approves every pack before it leaves the firm.

Two sentences were added during the week: the one about slide 2, on Tuesday, and the one about the age of the bank file, on Friday.

Exhibit 4: Sana's skill

Month-end memo. Use this when asked for a month-end memo.

  1. The first line is the status. Write READY when reconciled cash agrees with the cash balance in the ledger and nothing is unrecorded. Write NOT READY when anything is unrecorded.
  2. A timing item is in the ledger and not yet at the bank: a deposit the bank has not shown yet, or a check it has not paid yet. Count them. Do not list them unless asked.
  3. An unrecorded item is at the bank and not in the ledger: a returned check, a bank fee the company has not entered yet, a payment nobody entered. List each one directly under the status, with its date and amount.
  4. Then give reconciled cash, and how far it is above or below any minimum the client must keep.
  5. Never write that the books are correct. Write what was compared, and what agreed.

Exhibit 5: The request and the run note

The request, Tuesday, 10:05 a.m.

"Prepare the September 2027 month-end pack."

The run note, Tuesday, 10:30 a.m. (illustrative)

Files readBank lines, downloaded October 1. Ledger, downloaded October 1.
Lines3,395 bank lines, 3,404 ledger lines
Matched3,395 pairs
Left over9 ledger lines: 2 deposits, $6,120.00, and 7 checks, $9,300.00
Needs a look2 pairs with the same amount on the same day. Hamza confirmed them by hand.
Reconciled cash$198,300.00. Agrees with the ledger.
StatusREADY
Pieces updatedThe sheet, the memo, the slides, the cover

The run note, Wednesday, 10:00 a.m. (illustrative)

Files readBank lines, downloaded October 6. Ledger, downloaded October 1.
Lines3,396 bank lines, 3,404 ledger lines
Matched3,395 pairs
Left over9 ledger lines, as before. 1 bank line: a returned check, Erie Street Diner, $3,450.00, dated September 30.
Reconciled cash$194,850.00. The ledger says $198,300.00.
StatusNOT READY, 1 unrecorded item
Pieces updatedThe sheet, the memo, the slides, the cover
Changed since the last runThe bank's closing balance fell by $3,450.00. Reconciled cash fell by $3,450.00.

Exhibit 6: Two runs

Tuesday's runWednesday's run
Bank's closing balance$201,480.00$198,030.00
Add: deposits the bank has not shown yet$6,120.00$6,120.00
Subtract: checks the bank has not paid yet$9,300.00$9,300.00
Reconciled cash$198,300.00$194,850.00
The bank's minimum$195,000.00$195,000.00
Above or below the minimum$3,300.00 above$150.00 below
StatusREADYNOT READY, 1 unrecorded item
Approved byNisha, 4:10 p.m.Sana, 11:15 a.m.
Sent to Dana4:25 p.m.11:30 a.m., "this replaces Tuesday's pack"

Exhibit 7: The week

Cleveland is nine hours behind Karachi in October.

When, in KarachiIn ClevelandWhat happened
Fri, Oct 1, 4:00 p.m.7:00 a.m.The worker downloads September's bank lines, as it does every Friday
Mon, Oct 4, afternoonearly morningThe firm builds the setup
Tue, Oct 5, 10:05 a.m.1:05 a.m.Hamza's request. The run ends at 10:30 a.m.
Tue, 2:00 p.m.5:00 a.m.Sana's review. One rule is added, and the job runs again.
Tue, 4:10 p.m.7:10 a.m.Nisha approves $198,300.00
Tue, 4:25 p.m.7:25 a.m.The pack goes to Dana
Wed, Oct 6, 9:00 a.m.midnightThe morning check reports the returned check
Wed, 9:40 a.m.12:40 a.m.The second run. It ends at 10:00 a.m.
Wed, 11:15 a.m.2:15 a.m.Sana approves $194,850.00
Wed, 11:30 a.m.2:30 a.m.The new pack goes to Dana
Wed, 5:00 p.m.8:00 a.m.Dana reads it, and later calls her bank
Thu, Oct 7, 7:00 p.m.10:00 a.m.Dana meets her bank, with one pack
Fri, Oct 8, 6:00 p.m.9:00 a.m.Walter Brennan calls Nisha

Exhibit 8: By hand, and with the setup

All times are illustrative.

By handWith the setup
Download the files1 hourThe worker does it
Match about 3,400 lines8 hoursPart of a 25-minute run
Work out the figure1 hourPart of the run
Write the memo2 hoursPart of the run
Make the slides and the cover2 hoursPart of the run
Check and review3 hoursHamza 1 hour, Sana 45 minutes
Approve15 minutes15 minutes
People's time each monthAbout 17 hoursAbout 2 hours
Built onceNothingAbout 6 hours: three people, one afternoon
A late changeMost of a dayA second run, and about an hour of checking

Exhibit 9: Hamza's four tests

None of them uses Claude.

TestHowWhat it can miss
The arithmeticA calculator, three linesA wrong closing balance. On Tuesday Hamza took it from Claude's run note, not from the bank's statement, and missed the returned check. From now on, he compares the two first.
The matchingTwenty pairs, chosen at random, read against the bank's own fileA wrong pair among the other 3,375
The leftoversEvery leftover line, found in the ledgerA line the program wrongly matched, so that it is not left over
The piecesOpen the sheet, the memo and slide 2. Do they say the same?A figure in a place he did not open

Exhibit 10: A document editor and a workplace

A document editor, with AI help (Google's, or Claude's own docs)The firm's setup
Good forWriting and editing a document togetherA job that starts with files and ends with documents
Where the work startsWhen a person brings the figureAt the bank's file
Who does the matchingA person, perhaps with AI help in a spreadsheetClaude, with a program on a borrowed desk
Where the rules areIn people's heads, or pasted each timeIn the project and the skill
Next monthStart againRun it again
A late changeFix each file by handRun it again
History of a documentGoogle's editors keep a version history11Sheets have one, and the Docs page speaks of one.12,13 Anthropic's pages describe none for decks.14
Who is responsiblePeoplePeople

This table is fair only if you read it both ways. A firm can also build a job like this around other tools. We did not test Google's AI features for this case. The point is not which editor is better. It is the difference between a place to write and a setup that does the job.

Exhibit 11: Three answers for the owner

All figures are made up. Twenty packs a month: four companies, each with one month-end pack and four weekly packs.

1. Charge for the hours2. Charge for the pack3. Charge for the service
The priceAbout $110.00 a month-end pack and $40.00 a weekly pack$900.00 a month-end pack and $1,500.00 a month for the weekly workAbout $1,200.00 a month for each company
For each company, each month$270.00$2,400.00$1,200.00
For all four, each month$1,080.00$9,600.00$4,800.00
Today, for Lakeshore alone$2,400.00$2,400.00$2,400.00
What the client pays forPeople's timeA checked pack with a name on itThe packs, the morning check and corrections
Good for the clientCheap, and easy to understandA fixed price for a resultOne fee, and no bill for a bad month
Risk for the firmIt earns less each time it gets better. Nobody pays for building the setup.It must explain what the name is worthIt must decide how many clients three people can answer for
What it still trustsThat hours measure the workThat the client values the testsThat bad months are rare

And the line Nisha drew, for how far a task may run alone.

The setup mayA person must
Prepare onlyRead, match, build, post "ready for review"Check, approve, send
Send if READYAlso send, when the status is READY and the figure is above the minimumReview every NOT READY pack, and a sample of the others
Send alwaysPrepare and send every packRead what went out, afterward

Exhibit 12: Before you trust a setup

  1. Can you say what each part is for? One line for each: the project, the skill, the files, the schedule, the artifacts.
  2. Where does each file live, and which jobs can reach it? A job that must run without you cannot rely on a folder on your computer: it can be read only while the desktop app is open.
  3. What lasts after the desk is cleared? The method, the matched list and the run note must be saved somewhere else.
  4. Which rules are sentences, and which are settings? For example, which connected apps are turned on is a setting. A sentence asks. A setting stops. Know which rules have only a sentence.
  5. Does every run update the same pieces? Give the run their links, and check that it did not make new ones.
  6. Do the pieces agree? No piece reads another.15 Open each one.
  7. Who checks, and without what? At least one check must not use the AI.
  8. Who approves, by name, and what does the approval name? The pack and its figure.
  9. What does the morning check do when it finds something? Reports, and changes nothing, is the safe answer.
  10. How many times has it run? A setup that has run three times has not been tested by a strange month yet.
  11. Who may change the method, and who reviews the change? The instructions and the skill are the method. Keep a record of each change: who made it, when, and who agreed. 15.3 GitHub and the pull request shows one way.

Exhibit 13: Claude and ChatGPT side by side

Lakeshore works in ChatGPT. The same setup can be built there.

The needIn ClaudeIn ChatGPT
Rules for one piece of workA project's instructions6A project's instructions16
A methodA skillA skill, called with @ and its name
Files on your computerA folder, through the desktop appSource folders, through the desktop app
Connected appsConnectorsPlugins
Work at a set timeA scheduled task,10 or a routine in Claude in SlackA scheduled task17
The resultArtifacts: a sheet, a doc, a deck, a designIts own tools for the same jobs, as Build Something shows

The ChatGPT names come from ChatGPT 2.0, Set up the apps and Build Something. The names differ, and so do the plans that include each part. The four rules in The case at a glance are the same.

  1. The practical run sheet says how to run the exercise and record it. It has not yet been run.↑
  2. As verified on November 9, 2026, Anthropic says that in cloud sessions "the agent loop and code execution run in an isolated, temporary sandbox on Anthropic-managed infrastructure", and that each session's sandbox is "created when the session starts and destroyed when it ends". Claude Cowork architecture overview, Anthropic.↑↑2
  1. As verified on November 10, 2026, Anthropic says a routine in Claude Tag "uses that channel's connections with the same permissions as a typed request." Case 3 describes the worker and its channel. Set up routines, Claude Tag docs, Anthropic.↑↑2
  2. As verified on November 12, 2026, Anthropic says that on Team and Enterprise plans "an Owner or Primary Owner needs to enable" connectors "for the organization" before members can use them, and that each person still signs in. Owners can also set each connector's actions to Always allow, Needs approval or Blocked, and members cannot override this. Connectors belong to each person's account, not to a project. Use connectors to extend Claude's capabilities, Anthropic.↑
  3. As verified on November 11, 2026, Anthropic says projects "are available to all users, including those with free Claude accounts", that free users can make at most five, and that each project can have its own instructions and files. What are projects?, Anthropic.↑↑2
  4. As verified on November 11, 2026, Anthropic says skills "are available for users on Free, Pro, Max, Ultra, Team, and Enterprise plans" and that this "requires code execution to be enabled". How to create custom skills, Anthropic.↑
  5. As verified on November 11, 2026, Docs, Slides and Design are on every Claude plan, and Sheets needs a paid plan. What are artifacts and how do I use them?, Anthropic.↑
  6. As verified on November 10, 2026, Anthropic says "Claude Tag is available on Team and Enterprise plans" and "isn't available on individual plans (Free, Pro, Max, or Ultra)". It was in public beta. Work with Claude Tag, Anthropic.↑
  7. As verified on November 11, 2026, Anthropic says scheduled tasks "are available on Pro, Max, Team, and Enterprise plans", that each run opens as a new session, and that Claude "can use an included connector during a run, including writes". Automate work with scheduled tasks, Anthropic.↑↑2
  8. As verified on November 12, 2026, Google says "You can see changes that have been made to a document in Google Docs, Sheets, Slides, or Vids." A person with edit access can open an earlier version, name it and restore it. Find what's changed in a file, Google Docs Editors Help.↑
  9. As verified on November 12, 2026, Anthropic says "Version history shows earlier versions of a sheet and can restore an earlier version", "which also removes edits made after it." Get started with Claude Sheets, Anthropic.↑
  10. As verified on November 12, 2026, Anthropic's Docs page refers to a doc's version history. We have not tested what it keeps or for how long. Get started with Claude Docs, Anthropic.↑
  11. As verified on November 12, 2026, Anthropic says "Decks export to PowerPoint and PDF, or go straight to Google Slides as editable files." The help pages we checked describe no version history for decks. What are artifacts and how do I use them?, Anthropic.↑
  12. As verified on November 12, 2026, Anthropic says "Sheets can't pull live data from connected apps, and docs, decks, and dashboards can't use a sheet as a data source." Get started with Claude Sheets, Anthropic.↑
  13. As verified on November 11, 2026, OpenAI says project sharing "is available on ChatGPT Free, Go, Plus, Pro, Business, Enterprise, and Edu", and that "Project instructions apply only within that project". Projects in ChatGPT, OpenAI.↑
  14. As verified on November 11, 2026, OpenAI says "Scheduled tasks can use supported apps, including Gmail, Slack, and GitHub, when those apps are available for your account or workspace." Scheduled tasks in ChatGPT, OpenAI.↑
Owner
Panaversity