THE SUPERVISOR CARD · the run review · eight flags, every one with its lines
WAKES WHEN a multi-agent system, an agent crew or an orchestrator and its workers has run and someone asks what went wrong, who burned the budget, or why work stalled or was done twice
THE METHOD
template --out example — an example run (five agents of Example Co, fiction) and its budgets, to show the shape.
review run.jsonl --budgets budgets.json --out review — the flags, each agent's numbers, the unfinished items, RUN-REVIEW.md and review.json.
rules — the eight flags, their formulas and what to change.
The team changes the harness and reviews the next run.
REFUSES
A flag with no log line behind it
Dropping a flag
Pricing cost units
Judging whether the work itself was right
IT MAKES template — an example run log and budgets file (fiction) to see the shape; review — flag the run and write the run review; rules — the eight flags, their formulas and what to change.
THE LAW A flag stands on its log lines. No line, no flag; a flag the machine raised is never dropped. Cost stays in the log's own units; the card never prices anything. Every threshold is printed with the review, with where it came from. An item marked done and handed on is open again until its new owner finishes it. It reads only what the log recorded, and it reviews the run, not the work.
IN THE BOX scripts/supervisor.py · the flag rules · the method by hand · the sources note · ten tests.
PROOF selftest 26/26; a five-agent month-end run reviewed with nine flags and three unfinished items, and the trigger test, on the record (WITNESS.md) · witnessed 2026-10-13 · EDITION 1.
THE SEAL 96 of 100 · LEGENDARY · WITNESSED · Trigger 5 · Machinery 5 · Law 5 · Portability 5 · Proof 4
IT CAN BE WRONG an AI reading a text file: it can be wrong and can invent a number, a rule, a name or a date; check what matters against your own records and the source; the risk is yours. LICENSE-NOTE.md rides in the packet and is the licence.
THE PUNCH · SHA-256 OF SKILL.MD · 8D1FE53B88E3C5CFAE13A003
CITE THIS CARDTHE SUPERVISOR CARD, Edition 1. Lund Studio LLC, Westminster, Colorado. https://lundr.tech/?view=supervisor-card. Punch 8D1FE53B88E3C5CFAE13A003. Retrieved 2026-10-13.
What it is.
A supervisor's review of a run where several AI agents worked together: an orchestrator and its workers, a crew, a chain of subagents. From the run's log it shows who did what and used what, and raises the eight things that go wrong when agents work together:
Use it whenever a multi-agent system, an agent crew or an orchestrator and its workers has run and someone asks what went wrong, who burned the budget, or why work stalled or was done twice. Not for one agent's trace (THE TRACE CARD), splitting work before a run (THE CREW CARD), a receipt of what a run touched (THE RUN RECEIPT).
It carries a machine that shows its working: template — an example run log and budgets file (fiction) to see the shape; review — flag the run and write the run review; rules — the eight flags, their formulas and what to change — 26 checks in its selftest.
It answers requests such as “our agent crew ran overnight and the report never came out: what went wrong?”, “which agent burned the budget?”, “why did the reviewer and the writer keep passing the same task back?”, “give me a supervisor's report on last night's multi-agent run”.
Five laws, written into the card.
They are in the card itself, so the AI follows them.
A FLAG STANDS ON ITS LOG LINESA flag stands on its log lines. No line, no flag; a flag the machine raised is never dropped.
COST STAYS IN THE LOG'S OWN UNITSCost stays in the log's own units; the card never prices anything.
EVERY THRESHOLD IS PRINTED WITH THE REVIEWEvery threshold is printed with the review, with where it came from.
AN ITEM MARKED DONE AND HANDED…An item marked done and handed on is open again until its new owner finishes it.
IT READS ONLY WHAT THE LOG RECORDEDIt reads only what the log recorded, and it reviews the run, not the work.
The method.
template --out example — an example run (five agents of Example Co, fiction) and its budgets, to show the shape.
review run.jsonl --budgets budgets.json --out review — the flags, each agent's numbers, the unfinished items, RUN-REVIEW.md and review.json.
rules — the eight flags, their formulas and what to change.
The team changes the harness and reviews the next run.
What rides in the card.
In the card’s own voice, with a machine that shows its working.
THE REFERENCESreferences/SOURCES.md · THE-FLAGS.md · THE-METHOD.md
THE MACHINEscripts/supervisor.py · three commandstemplate — an example run log and budgets file (fiction) to see the shape; review — flag the run and write the run review; rules — the eight flags, their formulas and what to change. 26 checks in its selftest. Standard library only; nothing opens a network connection.
What it does once your AI has it.
template An example run log and budgets file (fiction) to see the shape
review Flag the run and write the run review
rules The eight flags, their formulas and what to change
On the record Example Co's five-agent month-end run reviewed: nine flags with their lines and three unfinished items; 44 of 48 in the trigger test
What it refuses, and where it hands off.
It refuses a flag with no log line behind it; dropping a flag; pricing cost units; judging whether the work itself was right.
one agent's trace, step by step → THE TRACE CARD
splitting work into briefs before a run → THE CREW CARD
a receipt of what a run read, changed and sent → THE RUN RECEIPT
the note an agent leaves for the person who takes over → THE HANDBACK CARD
Reads only what the log recorded; reviews the run, not whether the work was right; never prices cost units. It is an AI reading a text file: it can be wrong, so check what matters against your own records and the source.
What’s in the buy.
CARD.md — the whole card: its boot, law, method, machine, seals, refusals and hand-offs, ledger, score and what it carries · SKILL.md — the door your AI reads first · START-HERE.md — the first five minutes, for you
SOURCES.md + THE-FLAGS.md + THE-METHOD.md — in references/, in the card's own voice
scripts/supervisor.py — the machine: template · review · rules; standard library only, no network; selftest 26/26
WITNESS.md — the run on a realistic case as printed, and the trigger test on Haiku, Sonnet and Opus · SELFTEST.md — the machine's own test, as run at the cut · AUDIT.md, the twelve checks · MANIFEST.md, every file hashed
tests/evals.json — the cases it must take, hand off and refuse
MY-LEDGER.md — the card's memory, on your own machine · TIN.md, the nine lines on the tin · LICENSE-NOTE.md · HOSTS.md, where it runs
system/short-prompt.md and system/system-prompt.md — the doors for any AI without skills
PROOF · SELFTEST 26/26 · WITNESSED ON A RUN · TRIGGER TEST 44 OF 48 ON ITS OWN PROMPTS · HAIKU 12/16 · SONNET 16/16 · OPUS 16/16 · 0 WRONG PICKS · EDITION 1 · PUNCH 8D1FE53B88E3C5CFAE13A003
ON EVERY DOOR · A CARD FOR YOUR AI, A FILE FOR YOUR COMPUTER
In Claude: turn on code execution under Settings › Capabilities, then Customize › Skills › + › Create skill, and upload the zip as it is — START-HERE.md in the box walks you through it, and tells you how to keep the ledger. In ChatGPT, Gemini and the rest: a Project, a GPT or a Gem with the card’s files attached and the short door in the box pasted as its instructions. The words run anywhere an AI reads Markdown; the machines run wherever the host runs Python. Witnessed in Claude only. The host marks →
Take it.
Nº 970 — LAUNCHING SOON
Sold as is under the term printed. A card instructs an AI you operate; results depend on your model, your files and your judgement. Not medical, legal, tax or financial advice, never the person’s boss, and not a licence to practise. The terms →
One payment. The file is yours to keep and keeps working at its edition; new editions issued in the year after you buy reach you by email as replacement files. Tax is added at checkout where it applies. Nothing auto-renews.
A TERM, NOT A SUBSCRIPTION: NOTHING RENEWS ITSELF AND NOBODY IS CHARGED AGAIN. WHEN A TERM ENDS THE FILES STAY ON YOUR COMPUTER AND STAY READABLE — FREEZE, NOT LOSS.
THE MACHINE · IN THE OPEN · FROM THE CARD’S OWN WITNESS.MD AND SELFTEST.MD
supervisor.py, every command showing its working.
A small Python program. Your AI runs it; you can too. The blocks below are pasted from the card’s own record, as printed.
template
an example run log and budgets file (fiction) to see the shape
review
flag the run and write the run review
rules
the eight flags, their formulas and what to change
selftest
26 checks that prove the machine; it must print 26/26
RUN REVIEW · close-run.jsonl · 34 steps · 5 agents · 28 minutes
VERDICT · 9 flags · LOOP 1 · PING-PONG 1 · OVER BUDGET 1 · IDLE 1 · STUCK 1 · NO ESCALATION 3 · NO OWNER 1
FLAGS · with the log lines that prove them
LOOP collector lines 8,10,12,13 collector repeated 'fetch_statement' on BANK-1 4 times in a row
PING-PONG reconciler and reviewer lines 22,24,26,28,30 AP-3 went between reconciler and reviewer 5 times
OVER BUDGET collector lines 13 collector tokens 16,250 against a budget of 12,000 (135%), crossed at line 13
IDLE reporter lines 11,33 reporter did nothing for 22 minutes (lines 11 to 33) while other agents worked
STUCK reporter lines 33 reporter was still waiting when the run ended (1 minute after line 33)
NO ESCALATION collector lines 13 collector hit a budget breach at line 13 and nothing went to a person
NO ESCALATION reconciler lines 21 reconciler hit 3 errors in a row at line 21 and nothing went to a person
NO ESCALATION reporter lines 33 reporter hit a wait that never ended at line 33 and nothing went to a person
NO OWNER - lines 34 ACC-9 was worked with no owner (no assign, claim or hand-off before line 34)
AGENTS · steps · tokens · cost units · share · errors · last
collector 9 16,250 6.80 41.9% 0 handoff ok
orchestrator 3 1,200 0.60 3.7% 0 assign ok
reconciler 11 12,750 5.20 32.0% 3 handoff ok
reporter 4 2,050 0.79 4.9% 0 wait waiting
reviewer 7 7,000 2.85 17.5% 0 review ok
the run 34 39,250 16.24
UNFINISHED · BANK-1 (reconciler), RPT-1 (reporter), ACC-9 (no owner)
CHANGE
LOOP Give the orchestrator a repeat guard: after 3 identical steps, stop the agent and escalate. A loop usually means a result is not reaching the agent; check the tool's output first.
[… 12 more lines, trimmed here]
sed -n '/## The work items/,/## How this was decided/p' review/RUN-REVIEW.md
## The work items
| item | owner at the end | done |
|---|---|---|
| BANK-1 | reconciler | no |
| AP-3 | reviewer | yes |
| RPT-1 | reporter | no |
| ACC-9 | no one | no |
## How this was decided
THE SELFTEST, AS PRINTED
PASS three identical steps are a LOOP, with their lines
PASS the same action with different args is work, not a loop
PASS another agent's step between repeats does not hide a loop
PASS an item bounced four times between two agents is PING-PONG
PASS an agent past its token budget is OVER BUDGET, and with no escalation after it, NO ESCALATION
PASS the default budget '*' applies to every agent; an escalation after the breach clears NO ESCALATION
PASS the whole run past its budget is flagged on its own
PASS 23 minutes with no step while another agent worked is IDLE
PASS an agent listed as quiet (an orchestrator, a scheduler) is not called idle
PASS an agent still waiting at the end is STUCK, and unescalated
PASS three errors in a row with no escalation is flagged
PASS an orchestrator's escalation about the agent clears it
PASS work on an item no one owns is NO OWNER
PASS a second agent working an owned item is TWO OWNERS
PASS a hand-off moves ownership: no flag
PASS an item marked done and then handed on is open again until its new owner finishes it
PASS an item released and never taken again is NO OWNER
PASS a clean run says CLEAN with nothing unfinished
PASS per-agent totals add up to the run; the cost share is computed
PASS ISO times with Z or an offset are read
PASS a log mixing ISO times and seconds is refused
PASS template then review from an empty folder: flags found, exit 1, the same bytes twice
PASS the example run shows all eight flag kinds, and its quiet orchestrator is not called idle
PASS RUN-REVIEW.md carries the flags, what to change, the rule, the source and the limit
PASS a broken line or a missing field exits 2 naming the line
PASS rules prints all eight flags with their changes
selftest 26/26
THE TRIGGER TEST. The card's line was tried on Haiku, Sonnet and Opus beside 16 neighbouring skills, in a clean environment: 44/48 on its own prompts (Haiku 12/16 · Sonnet 16/16 · Opus 16/16), 0 wrong picks; no skill (answered directly) 4.
THE RUNS ARE PASTED FROM THE CARD’S OWN RECORD AS PRINTED. THE CASES ARE THE STUDIO’S OWN EXAMPLES (EXAMPLE CO IS FICTION); NO PRIVATE PERSON IS NAMED.
QUESTIONS AT THE TABLE
Asked before the wire.
What does it need from me?
Python 3; the run's log as JSON Lines, and its budgets if the team has them.
Does it send anything anywhere?
No. The machine is a standard-library Python program that runs on your own computer and sends nothing anywhere; the words run wherever an AI reads Markdown. No server, no account, no key.
What is on the record?
The selftest (26/26), a run on a realistic case with every command and verdict pasted into WITNESS.md as printed (Example Co's five-agent month-end run reviewed: nine flags with their lines and three unfinished items; 44 of 48 in the trigger test); and the trigger test, the card's line tried on Haiku, Sonnet and Opus beside 16 neighbouring skills: 44/48 on its own prompts, 0 wrong picks; the twelve-check audit (AUDIT.md). The punch on the face is the SHA-256 of SKILL.md, so a downloaded copy can be checked against this room.
Can it be wrong?
Yes — it is an AI reading a text file. Reads only what the log recorded; reviews the run, not whether the work was right; never prices cost units.
What will it never do?
It refuses a flag with no log line behind it; dropping a flag; pricing cost units; judging whether the work itself was right.
Who may use it?
One person, any number of their own AIs and computers; a team needs a company licence.
