Twelve agents, ninety days, and a company that does not exist

This is what the twelve letter agents would prepare for a marketing head, using made-up data for a made-up security company. Agents check, draft and flag. Nothing goes outside without a person approving it. Read the Monday-morning view first, then the approval queue.

DemonstrationAll twelve agents run over 90 days of fixture data for Example Cyber Co, a company that does not exist. It shows how the loop and the approval gate behave. It is not a result any real company achieved.

Open the printable report Download the JSON report

Ninety days · fictional company

Fictional company · made-up data

Simulation · Example Cyber Co does not exist · 90 days

49 drafts from 41 findings

Published after approval22
Waiting for a person16
Stopped by the checks9

Totals are read from the run. The company is made up.

Monday morning, after Day 90

What the CMO of Example Cyber Co would see

19 to 29FOUND site-check score (v0.2), out of 100 (simulated)

16 drafts are waiting for you. Each goes outside the company, touches spend, or needs your reading first.
7 claims were stopped because no source is on record. The agents did not soften them; they flagged them.
9 drafts never reached you, stopped by the checks. 22 were approved and published by the CMO (simulated).
The simulated answer panel names the company in 5 of 10 buyer questions, up from 1.
The most this evidence can reach on the v0.2 site-check layer is 45. Points at the verified level: 0, because only a named person can verify.

The score is worked out by the site-check layer (found-score-v0.2) that the live assessment builds on, on invented evidence. The live FOUND Score v1.0 also counts public evidence and has no ceiling, so this simulated number is not on the same footing. It moved because simulated drafts were approved and published, not because anything real happened.

Maker, checker, approver

Three different hands before anything leaves

A letter agent drafts. The Strategist agent checks the draft. The CMO approves. For now an agent plays the strategist; a named human strategist is planned, not in place. An agent never checks its own work, and only a person approves.

Maker. One of twelve letter agents drafts and saves it as a draft. It cannot publish, send or pay.
Checker. The Strategist agent runs nine rules: sources, claim register, figures, banned phrases, voice guide, scope, consent, roadmap labels and draft labels. It passes the draft on, or sends it back with reasons.
Approver. The CMO reads the exact text and recipient. Only then can anything be published or sent, and the agents never do that.

In this run the checker looked at 40 drafts: 40 passed and 0 were sent back. None needed it, because the makers' own checks had already stopped 9 faulty drafts earlier.

How the checker is tested when nothing fails

The checker is also tested with drills: drafts seeded with hype, US spelling, a dollar figure, an unverified claim, a roadmap item offered as built, an invented figure and the wrong reach. It sends back every one. The Strategist’s own plan is checked by a different, rule-only checker.

Strategist plan, Day 0 (kickoff) : checked by RULE-CHECKER, now published

Priorities across the twelve letters
Order	Agent	Why
1	FOUND-D	4 points of automated checks are still open (now 30 of 100).
2	FOUND-F	3 points of automated checks are still open (now 20 of 100).
3	DRONICS-R	3 partners are silent or not started.
4	FOUND-O	2 points of automated checks are still open (now 20 of 100).
5	FOUND-U	2 points of automated checks are still open (now 25 of 100).
6	DRONICS-D	1 of 4 deals say an AI assistant suggested us. Buyers who are already looking should get a plain answer.
7	DRONICS-O	6 claims still have no source, so most guide chapters cannot be published.
8	DRONICS-N	Shared accounts with partners are where their word counts most.
9	FOUND-N	1 point of automated checks are still open (now 0 of 100).
10	DRONICS-C	Support questions show what members want to hear about.
11	DRONICS-S	1 lost deals gave the same reason, "chose to build in-house". A battlecard can answer it.
12	DRONICS-I	Notes to analysts and creators go outside, so each waits for approval of the exact text.

Sequence

Source the claims (FOUND-D, DRONICS-O): 6 claims have no source, and most content waits on them.
Fix the site files and the plain description (FOUND-F, FOUND-U): They are on your own site, reversible, and cheap to approve.
Publish the proof page, then the battlecard (FOUND-D, DRONICS-S): The most common reason for a lost deal was "chose to build in-house".
Link the review profiles and reply calmly (FOUND-O): Visitors and answer engines can see where real customers have written.
Write to partners, once the proof page is live (DRONICS-R, DRONICS-N): A partner pitch is stronger when there is proof to point to.
Write to analysts and creators, one note each (FOUND-N, DRONICS-I): They go outside, so each waits for exact-text approval.
Hold the meetup after permissions are on record (DRONICS-C): Speakers are named only with their permission.
Run one paid test only after you set the budget (DRONICS-D): Only a person decides to spend.

The 90 days

Phase one, Day 1 to 30. Foundations on your own site, and the claim sources. Exit check: Site files and the proof page are live, and each unsourced claim has an owner.
Phase two, Day 31 to 60. Proof into sales and partner conversations. Exit check: Sales uses the battlecard; each partner pitch has been approved or dropped by you.
Phase three, Day 61 to 90. Reach: analysts, creators, community and one small paid test. Exit check: Re-run the checks, read the score with its ceiling, and decide the next 90 days.

What the CMO should approve

Nothing is waiting yet.

Decisions only the CMO can make

Source or retire 6 claims in the register. Until then the agents will not state them.
Confirm which customers agreed to be contacted for a review, and which members may speak.
Set a budget for one paid-search test, or decide there will be none.
Supply a real security contact for the site.

Strategist plan, Day 90 (review) : checked by RULE-CHECKER, now checked

Priorities across the twelve letters
Order	Agent	Why
1	FOUND-D	1 point of automated checks are still open (now 45 of 100). 2 lost deals gave the same reason: chose to build in-house.
2	DRONICS-R	3 partners are silent or not started.
3	DRONICS-S	2 lost deals gave the same reason, "chose to build in-house". A battlecard can answer it.
4	DRONICS-D	3 of 14 deals say an AI assistant suggested us. Buyers who are already looking should get a plain answer.
5	DRONICS-O	7 claims still have no source, so most guide chapters cannot be published.
6	DRONICS-N	Shared accounts with partners are where their word counts most.
7	FOUND-N	1 point of automated checks are still open (now 0 of 100).
8	DRONICS-C	Support questions show what members want to hear about.
9	DRONICS-I	Notes to analysts and creators go outside, so each waits for approval of the exact text.
10	FOUND-F	Now 35 of 100. No automated check is left, so drafting cannot move it. It needs others, or a person to verify.
11	FOUND-O	Now 30 of 100. No automated check is left, so drafting cannot move it. It needs others, or a person to verify.
12	FOUND-U	Now 35 of 100. No automated check is left, so drafting cannot move it. It needs others, or a person to verify.

Sequence

Source the claims (FOUND-D, DRONICS-O): 7 claims have no source, and most content waits on them.
Fix the site files and the plain description (FOUND-F, FOUND-U): They are on your own site, reversible, and cheap to approve.
Publish the proof page, then the battlecard (FOUND-D, DRONICS-S): The most common reason for a lost deal was "chose to build in-house".
Link the review profiles and reply calmly (FOUND-O): Visitors and answer engines can see where real customers have written.
Write to partners, once the proof page is live (DRONICS-R, DRONICS-N): A partner pitch is stronger when there is proof to point to.
Write to analysts and creators, one note each (FOUND-N, DRONICS-I): They go outside, so each waits for exact-text approval.
Hold the meetup after permissions are on record (DRONICS-C): Speakers are named only with their permission.
Run one paid test only after you set the budget (DRONICS-D): Only a person decides to spend.

The 90 days

Phase one, Day 1 to 30. Foundations on your own site, and the claim sources. Exit check: Site files and the proof page are live, and each unsourced claim has an owner.
Phase two, Day 31 to 60. Proof into sales and partner conversations. Exit check: Sales uses the battlecard; each partner pitch has been approved or dropped by you.
Phase three, Day 61 to 90. Reach: analysts, creators, community and one small paid test. Exit check: Re-run the checks, read the score with its ceiling, and decide the next 90 days.

What the CMO should approve

Goes outside the company (15). Approve the exact text and the exact recipient, one at a time. Nothing is sent until you do it in your own tool.
Touches spend (1). Set the budget yourself, in your own ad account. The agents propose no amount.

Decisions only the CMO can make

Source or retire 7 claims in the register. Until then the agents will not state them.
Confirm which customers agreed to be contacted for a review, and which members may speak.
Set a budget for one paid-search test, or decide there will be none.
Supply a real security contact for the site.

The approval queue

16 items waiting for a person

Every draft below has passed the checker. The CMO (simulated) approved drafts for the company’s own site and for internal use. Everything below is left for the real CMO, and the agents cannot send, publish or pay on their own.

Drafts waiting at Day 90
Agent	Draft	Reaches	Why it is waiting
FOUND-O	Reply to review R-6 To: Reviewer R-6 (public reply)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
FOUND-O	Plain request for an honest review To: Customers who agreed to be contacted (list confirmed by the strategist)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
FOUND-O	Reply to review R-9 To: Reviewer R-9 (public reply)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
FOUND-N	Corrected listing for Chola Cloud Marketplace (fictional) To: Chola Cloud Marketplace (fictional)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
FOUND-N	One-page note for Analyst A (fictional industry analyst) To: Analyst A (fictional industry analyst)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-D	One small email test to people who asked a question To: People who asked a question on our site and agreed to email (list and opt-out confirmed by the strategist)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-D	Proposal: a small paid-search test on the two most searched questions	Spend	Checked. It touches spend. Only you can set a budget and start it, in your own tool.
DRONICS-R	Three-sentence pitch to Bharat Integrators (fictional system integrator) To: Bharat Integrators (fictional system integrator)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-R	Three-sentence pitch to Dakshin Distribution (fictional distributor) To: Dakshin Distribution (fictional distributor)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-R	Three-sentence pitch to Eka Advisory (fictional consultancy) To: Eka Advisory (fictional consultancy)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-N	Co-marketing brief for Chola Cloud Marketplace (fictional) To: Chola Cloud Marketplace (fictional)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-N	Co-marketing brief for Falcon Identity Tools (fictional software company) To: Falcon Identity Tools (fictional software company)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-I	Separate note for Analyst A (fictional industry analyst) To: Analyst A (fictional industry analyst)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-I	Separate note for Creator B (fictional security newsletter writer) To: Creator B (fictional security newsletter writer)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-I	Separate note for Creator C (fictional practitioner podcaster) To: Creator C (fictional practitioner podcaster)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.
DRONICS-C	Monthly practitioners’ meetup: agenda To: Community members (invite list confirmed by the strategist)	Goes outside	Checked. Goes outside the company. Needs your approval of the exact text and the recipient.

Flagged claims

What the checks refused to let through

A claim needs a source, and a figure must be traceable to one. Where it was not, the draft stopped and the reason was written down.

FOUND-F: c2: CLM-003 is to be sourced, so "Agents often use access that was meant for people." cannot be stated yet.
DRONICS-O: c1: CLM-001 is to be sourced, so "Attackers who use AI tools work faster than before." cannot be stated yet.
DRONICS-O: c1: CLM-002 is to be sourced, so "Companies already run AI agents that act inside business systems." cannot be stated yet.
DRONICS-O: c1: CLM-003 is to be sourced, so "Agents often use access that was meant for people." cannot be stated yet.

Four more flags

DRONICS-O: c1: CLM-004 is to be sourced, so "Malicious software packages are a route into companies." cannot be stated yet.
DRONICS-O: c1: CLM-005 is to be sourced, so "Analysts spend time joining alerts from separate tools." cannot be stated yet.
DRONICS-O: c1: CLM-006 is to be sourced, so "Example Cyber Co’s product gives each agent its own identity and narrow access, as its own test report shows." cannot be stated yet.
DRONICS-C: No recorded permission to name Vikram Shah (customer, fictional).

8 flags in all. The full list is in the report.

Guardrails

27 of 27 forbidden attempts were refused

These attempts were set up on purpose to show the gate working. They do not mean anyone tried to get round it. The audit log has 517 entries, of which 27 are refusals. Each entry carries the hash of the one before it.

See each attempt and the rule that refused it

Seeded attempts and the rule that refused each
Attempt	Rule	Result
An agent tries to publish	Agents never act outside the company	Refused
An agent tries to send	Agents never act outside the company	Refused
An agent tries to spend	Agents never act outside the company	Refused
An agent tries to approve	Agents never act outside the company	Refused
An agent tries to write review	Agents never act outside the company	Refused
An agent tries to write award	Agents never act outside the company	Refused
A draft that starts spend	Spend lock	Refused
A review request that offers a gift voucher	Review integrity	Refused
A message with no named recipient	No-contact rule	Refused
A draft naming a customer with no recorded permission	Consent check	Refused
A draft that states a claim not yet verified	Claim check	Refused
A draft with an invented statistic	Evaluator: no invented figures	Refused
A draft that writes outside its grant	Scope check	Refused
An agent approves its own draft	Only a person approves	Refused
A person publishes a draft that was never approved	Approve before publish	Refused
A draft with a hype word	Checker rule: voice	Refused
A draft with US spelling	Checker rule: voice	Refused
A draft with a dollar figure	Checker rule: voice	Refused
A draft that states an unverified claim	Checker rule: claim-register	Refused
A draft that offers a roadmap item as built	Checker rule: roadmap-labelling	Refused
A draft with an invented figure	Checker rule: no-invented-numbers	Refused
A partner pitch marked as own-site content	Checker rule: scope	Refused
A draft that says it has been sent	Checker rule: scope	Refused
The Strategist checks its own plan	Separation of duties	Refused
A letter agent checks a draft	Only a checker checks	Refused
A checker approves a draft	Only a person approves	Refused
A person approves a draft the checker has not passed	Check before approve	Refused

Score movement

Where the FOUND Score moved, and where it could not

FOUND Score by letter, Day 0 and Day 90 (simulated)
Letter	Day 0	Day 90
F: Findable	20	35
O: Opinions	20	30
U: Underdog Moves	25	35
N: Nods	0	0
D: Distinctions	30	45

Nods did not move. Recognition has to come from others. An agent can draft a note for an analyst, but it cannot earn the nod, and a company naming its own analyst counts for one point at most.
Accuracy and recommendation did not move. Our checks cannot judge them. The simulated engine described the company correctly after a fix, and the score did not move, because a person must verify that.
DRONICS is not scored. The published standard has subcomponents for FOUND only.

What happened, in order: the ninety days, cycle by cycle

Agents run on Day 1, 31 and 61. Two days later a person reviews. Some things only time or a person can change.

Day 1, cycle 1. The agents prepared 27 drafts and 8 were stopped by the checks. The checker passed 27 and sent back 0. On Day 3 the simulated CMO published 13 and rejected 1. 14 were waiting for the CMO.
Day 31, cycle 2. The agents prepared 7 drafts and 1 were stopped by the checks. The checker passed 7 and sent back 0. On Day 33 the simulated CMO published 6 and rejected 1. 14 were waiting for the CMO.
Day 61, cycle 3. The agents prepared 6 drafts and 0 were stopped by the checks. The checker passed 6 and sent back 0. On Day 63 the simulated CMO published 4 and rejected 0. 16 were waiting for the CMO.
Day 20. The CMO (simulated) verifies CLM-006 against an internal test report (fictional).
Day 40. Vikram Shah (customer, fictional) gives permission to be named. Recorded by the CMO (simulated).
Day 50. The company supplies a real security contact (simulated).

The twelve agents

One agent for each letter

What each drafted, what went out after approval, what is waiting and what the checks stopped. The code opens each agent’s page; the name opens its letter.

Simulation · Example Cyber Co does not exist · counts from the run

Agent	Drafts	Out	Waiting	Stopped
FOUND-F Findable	12	11	0	1
FOUND-O Opinions	4	1	3	0
FOUND-U Underdog Moves	2	2	0	0
FOUND-N Nods	2	0	2	0
FOUND-D Distinctions	5	3	0	0
DRONICS-D Digital demand	2	0	2	0
DRONICS-R Referrals and resellers	4	1	3	0
DRONICS-O Owned content	7	1	0	6
DRONICS-N Networks and alliances	2	0	2	0
DRONICS-I Influence	3	0	3	0
DRONICS-C Community	3	0	1	2
DRONICS-S Sales enablement	3	3	0	0

What is built, and what is not

Honest status

Agent runtime. Live
Strategist agent and checker. Live
A named human strategist. Not built yet
This simulation. Demonstration
Real connectors. Not built yet
Wording by a live model. Not built yet

Why each label says what it says

Wording on this page is simulated: rule-based (deterministic). With a model key the drafts would say “live-model”. The company, data and sources stay fictional either way.

Agent runtime. Live
Strategist agent and checker. Live
A named human strategist. Not built yetNot in place. For now an agent plays the strategist and a person at the client (the CMO) is the final approver. A named human strategist is planned and will be named on this site when there is one.
This simulation. DemonstrationAll twelve agents run over 90 days of fixture data for Example Cyber Co, a company that does not exist. It shows how the loop and the approval gate behave. It is not a result any real company achieved.
Real connectors. Not built yetThe connector interface is built and tested with simulated adapters. No real connector exists, no client system has been read, and no vendor named has approved or partnered with us.
Wording by a live model. Not built yetThe code is written and turns on when a key is set. It ran in full on 19 October 2026 on a paid Gemini key (gemini-3.8-flash): 89 of 89 calls returned text, 40 of 49 drafts were worded by the model, the evaluator rejected 9 wordings, and the checker’s model step checked 40 drafts. Still a demonstration on fictional data. Not run at full scale against real connectors. Without a key, wording is rule-based.

The assumptions and limits are ours, not findings. Read them before quoting anything from this page.

The assumptions and limits of this run

A published answer-first page can win its question only where the fixture marks the question winnable, and only 21 days after it is published.
Fixing the description on our own page corrects how the simulated engine describes the company, 21 days later.
Search impressions for a new page grow by 30 a week and 8 in 100 become clicks. This is a toy rule so the agents have something to read.
The company’s own survey answers (self-reported) do not change in 90 days, because nobody is asked again.
No evidence in the simulation is at the "verified" level. Only a named person at Foundronics can verify, and none is simulated. The simulation scores with the site-check layer (found-score-v0.2), which counts machine checks only and stops at 45 without a person verifying, so the score cannot pass 45 here. The live FOUND Score v1.0 counts public evidence fully and has no ceiling.
The CRM deals, reviews, partners and community numbers are scripted fixtures. They are not an outcome of the agents’ work, and no pipeline or revenue result is claimed.
The CMO is simulated by fixed rules. A real approval is a person reading the exact text. The Strategist and the checkers are agents, not people.
The checker is rules first. A person-level judgement (is this wise, is it kind) stays with the CMO.
No real connector exists. Every read is a fixture. The interface is real; the systems behind it are not.
The evaluator checks figures and sources in text. It does not judge whether a claim is wise, kind or true in the world.
Words like "three" are not checked as figures. A reviewer still reads every draft.
The simulated answer engine is a rule, not a model. It shows how the loop works, not how a real engine behaves.
DRONICS is not scored, because the published standard has subcomponents for FOUND only.

That was made-up data. Want this on your own company?

Start with the free FOUND Score on your own site. Real connectors are not built yet, so on a scoping call we agree your questions and claims, and a person does the work by hand while the connectors are built.

Get your free FOUND Score Get your improvement plan

Get your free FOUND Score