1 of 21

I built a support team of AI agents

so I could get my time back

BUILD 1�The job-search

system

$$$$�People

bought it

TIME�People needed

help

BUILD 2�A chat support

for the fix

Laid off in December, then built an AI job-search system and sold it.

Then supporting the buyers started eating the time it was supposed to give.

So I built a second system to carry the load.

Renee Romero . AI Systems Designer . 2026

2 of 21

SECTION 01

The problem

What support looked like before this build.

3 of 21

THE TURNING POINT

“My Claude doesn’t know what your Claude knows.”

A buyer said this on a support call. He suggested I build a custom GPT.

That was key.

The people who bought my job search system had a 64-page setup guide and to use their own AI.

The fix was NOT more documentation.

It had to be something that already carried the knowledge.

Renee Romero . AI Systems Designer . 2026

3

4 of 21

THE PROBLEM, IN NUMBERS

Supporting buyers ate my time

64 - page setup guide

9 → 64

pages the Setup Guide grew, and still did NOT close the gap

13

paying customers across 4 product-packages

Call after call

live support, each one straight out of deep work

The guide told buyers to upload the workflow files to their own LLM.

Fine on paper.

In practice the questions kept coming.

Renee Romero . AI Systems Designer . 2026

4

5 of 21

THE PLAN

What the support would need to do

I knew the problem, not the build.

So I brainstormed with Claude: what would this thing have to do to support any buyer, no matter how they set it up?

Know every workflow�by node name, not by guess

Know the Notion setup�field mappings and all

Know what the buyer owns�which package, what it includes

Fetch the answer�hand off cleanly when one agent can’t solve it

Say when it doesn’t know�instead of making something up

Cost less than my time�so most questions never reach me

Renee Romero . AI Systems Designer . 2026

5

6 of 21

SECTION 02

The system

13 agents, 1 production Worker, 4 KV namespaces.

7 of 21

THE ARCHITECTURE

The stack it runs on

Claude said multi-agent was the right shape and to use these parts to build it.

Cloudflare Workers

one production Worker, entry point and routing

Claude API

Sonnet 4.6 for most, Opus 4.6 for the hard cases

Gumroad webhook

a purchase creates the buyer token

Resend email

escalations to me plus the daily gap digest

TypeScript

one file, 4,378 lines, the entire system

Renee Romero . AI Systems Designer . 2026

7

8 of 21

THE MULTI-AGENT ARCHITECTURE

Multi-layered escalation

Renee Romero . AI Systems Designer . 2026

9

9 of 21

THE ARCHITECTURAL FLOW

How the pieces connect

Renee Romero . AI Systems Designer . 2026

8

10 of 21

WHAT THE BUYER SEES, 1 OF 2

The routing happens in the open

The buyer is on a hosted URL link in their browser.

They ask about WF08 cost.

The orchestrator routes checks with WF08 specialist.

1 Routing

2 Checking

Renee Romero . AI Systems Designer . 2026

10

11 of 21

WHAT THE BUYER SEES, 2 OF 2

What the buyer gets: pulled from the real files

WF08 specialist answers.

Renee Romero . AI Systems Designer . 2026

11

12 of 21

THE COST ARCHITECTURE

30 free, then 3 prepaid tiers

30 free / day

reset on a daily timer, never touches paid balance

$5 / $15 / $25

optional prepaid credit after the free tier, never expires

2 hard caps

a 30-message buyer gate and a workspace ceiling, nested

2 stores, never crossed

free count and paid balance live apart

Renee Romero . AI Systems Designer . 2026

12

13 of 21

THE BOUNDARIES IT ENFORCES

Scope and security: 2 things it refuses

Scope refusal�off-topic ask,

it declines and

points to a general assistant

Leak guard�asked for the workflow file,

it won’t hand back the paid product,

even to a buyer

Renee Romero . AI Systems Designer . 2026

13

14 of 21

LAUNCHED

Live on 05-28-2026

13 buyers, 1 production system.

Everything that follows happened after it went live, in front of real customers.

15 of 21

THE CATCH BEFORE LAUNCH

The data didn’t match the notes

Before I sent the email announcement to my buyers, I checked production directly.

I didn’t trust the handoff notes.

THE NOTES SAID

  • 13 customers loaded�Token store populated�Ready to send the announcement

PRODUCTION ACTUALLY HAD

  • 0 of 13 customers loaded�Test data and one real customer�The announcement would have sent 13 buyers to a 404

Verification caught a silent failure.

Renee Romero . AI Systems Designer . 2026

15

16 of 21

THE CATCH AFTER LAUNCH

The morning-after email digest

The digest flagged a 30-message session.

Claude checked the logs.

Found the specialist guessed its replies.

No red dot and no error notifications.

This = silent failure.

The fix

  • Not more documentation.
  • Grounding every specialist in the real files.
  • I grounded all 9 workflow specialists and validated each live on 05-30-2026.

Renee Romero . AI Systems Designer . 2026

16

17 of 21

SECTION 03

How it was built

5 weeks, a non-engineer, and 1 phase table.

18 of 21

HOW IT GOT BUILT

Claude frustrated me, so I built a table to track everything

I didn’t know how many agents this needed. I knew overloading one agent backfires, so it had to be split, but how many?

Early on I hit frustration. I’m not technical, and I couldn’t hold all the “not yet” pieces across chats that reset between sessions.

So we built a phase table that carried the plan forward, similar to lesson planning from my former career as a teacher.

PHASE & STATUS TABLE

A Auth + token lookup

Done

B Orchestrator routing

Done

C Specialist grounding

Done

D Hedge cascade

Half-wired

E Cost caps + tiers

Done

F Digest + gap logging

Half-wired

G Backfill + launch

Not started

H Buyer announcement email

Not started

Renee Romero . AI Systems Designer . 2026

18

19 of 21

BEFORE VS. AFTER

What changed

BEFORE

AFTER

Support

64-page PDF, ad-hoc calls, self-serve

on-demand chat, gated, knows the codebase

Time

15 min to 3.5 hrs per call, unpaid

~$0 of my time, asynchronous

Knowledge

whatever the buyer’s AI guessed

grounded in the real JSONs and schema

System

none, buyers recreated it

13 specialists hold the knowledge

Build effort 80+ hrs over ~5 weeks Live 13 buyers as of 05-28-2026

Renee Romero . AI Systems Designer . 2026

19

20 of 21

WHAT THIS BUILD TAUGHT ME

The takeaways

01

The user names the real problem, not the designer

The 64-page guide wasn’t the problem. The problem was the gap between where the knowledge lived and where the buyer needed it. The buyer told me. I listened.

02

Both failures pointed to human judgment in the loop

The records that weren’t loaded and the ungrounded answers were both caught by me, not the system. That is where the trust belongs.

03

This build added a new toolkit

The vocabulary: scope refusal, leak guard, grounding, hedge cascade, backfill. The stack: Cloudflare Workers, Resend, TypeScript, and a Gumroad webhook.

Renee Romero . AI Systems Designer . 2026

20

21 of 21

I’m an AI Systems Designer, not an engineer.

The skill is naming the real problem, then building only what solves it.

Specification Precision . Evaluation Design . Multi-Agent Decomposition . Failure Pattern Recognition�Trust and Security Design . Context Architecture . Cost and Token Economics

reneeromero326@gmail.com . linkedin.com/in/renee-romero . muralderomero.com