Agents at Work  ·  Sydney

Build,
don't rent

How the plant guy — no CS degree, no dev team — built the AI operating system his business actually runs on.

Craig Helmers — The AI OperatorOne system. Every team. Real work.
SpeakerOpen with the plant angle straight away — own it. "I put plants in offices for a living. Stay with me." This is a war story, not a tech talk. 25 minutes, then 15 of live demo.
01The breaking point
$25k

a year for admin support.
Nine months in — things were still getting missed.

The choice: retrain and re-onboard another human… or build a team that never forgets its training.

Act I — StakesI didn't do this because I love tech
SpeakerThe real opening line: "I didn't build an AI team because I love technology. I built one because I was paying $25K a year and things were STILL getting missed." No blame on the person — the model was broken: every departure resets training to zero. That's the pressure that started everything.
02The villain
RentingDemos that
impress
BuildingSystems that
work
Act I — Name the enemyScattered SaaS · AI theatre · zero change
SpeakerName the villain early: renting. Another subscription, another AI demo that wows for a week and changes nothing about how work moves. Every failure in this story is a renting failure. We kill it at the end.
03The bet

I wasn't technical.
I was stubborn.

Day one of GPT, I swapped my browser search bar for the model — before I knew what it was for. Curiosity became reps. Reps became an edge. Then one weekend and 1,000+ screenshots later, I had OpenClaw running in the guts of my business.

Act I — RepsWilling to look silly for a while
SpeakerKeep it humble. Tech-savvy, not a coder. The 1k screenshots weekend is the "I was in the mud" proof — you didn't rent this, you crawled through it.
04Where that got us
5

of us run the company
— two are human

300+

clients under management

$1.5m

recurring revenue

The rest of this talk is how — including the expensive, embarrassing parts they leave out of the demo videos.

Act I — The hookNow the room wants the how
SpeakerLand the "two are human" beat and pause — it reframes headcount forever. Then promise the failures up front: "$25K burned, two days of Xero cleanup, a rebuilt app. You're getting all of it."
05First wrong answer
What I didGive them
tools
What they neededGive them
context
Act II — Hammer, nail, smashed thumbTools made them powerful, not smart
SpeakerIf every agent has a hammer, every problem looks like a nail. Access to tools without context is how you automate mistakes at scale. This sets up the Xero disaster two slides from now.
06What context actually means

The kindergarten teacher

A kid asks what a word on page 30 means. The teacher reads the entire book — then comes back to page 30 to answer.

That's most AI setups: drowning in everything, slow and expensive, to answer one thing.

Most systems

Read everything, then answer

Ours

Go to page 30 first.
Read more only if you must.

Act II — Context disciplineRelevance beats volume
SpeakerThis is the whole context-engineering discipline in one image, and it's yours — tell it warm, like a story about an actual kid. Payoff: playbooks and context caches exist so agents open the right page instead of re-reading the library. Faster, cheaper, and — surprisingly — smarter.
07The game changer

Workflow… or agent?

A workflow follows steps. An agent makes calls. Most "AI teams" are workflows in a trench coat. The difference is four things:

T

Tools

What it can do.

C

Context

What matters right now — playbooks, client history, source truth.

Q

Questions

What it asks before acting.

J

Judgement

What best serves the North Star.

Act II — TCQJThe whole model on one slide
SpeakerThis was THE lightbulb: stop asking "what should it do" and start asking "what does it need to decide well." Tools, Context, Questions, Judgement — miss one and you've built a workflow that lies about being an agent.
08Confession № 1

Two bots.
My accounts.
No onboarding.

I let two agents loose on Xero like they'd read the manual. They hadn't.

✕Double-entered bills
✕Invented account codes
✕Made things up with total confidence
=Two full days of cleanup
Act II — The embarrassing bitYou onboard AI like you onboard staff
SpeakerTell it like a confession, get the laugh, then the lesson: I'd never throw a new hire at the accounts with no induction — I did exactly that to two bots. Today: same job runs with the invoices found, prepared and coded — and a human approving. Human in the loop isn't a weakness, it's the design.
09Confession № 2

The app that lost a week

My first field-service build dropped a week of plant swaps and missed rooms. I went back to staff and re-entered it by hand. Five weeks in, I tore it down and rebuilt — because I'd built rooms before scaffolding.

Then

Features first, foundations never

Now

A map per client. Pass / fail per visit. Nothing silently lost.

Act II — Scaffolding beats speedGot carried away building, forgot to check
SpeakerSecond confession: got carried away adding features to every tab, never got in the weeds to verify. The rebuild lesson became house law: research → challenge → spec → build → QA. Nothing ships because the builder says so.
10The law

What cannot
be shown
was not done.

Agents don't get to say "done." Every claim needs an artifact — a file, a screenshot, a test result. And the builder never grades his own work.

Act II — Trust, solvedWe don't trust it. We don't have to.
SpeakerThe room's silent question is "how do you trust it?" Answer slowly: we don't. We built a system where trust isn't required — evidence is. This law came directly out of the two confessions you just heard. Every fix since has been "show me," never "trust me."
11Breakthrough

North Star beats instructions

Rules stop bad moves. A North Star helps an agent choose good ones — and we hang one over everything: every agent, every playbook, every build.

?

Mission

What are we really trying to achieve — beyond this one task?

✕

Boundary

What must never break, no matter how clever the shortcut looks?

✓

Proof

How do we know it worked? Evidence, not vibes.

Act II — The spineMission · Boundary · Proof
SpeakerPlain example: preserve client trust (mission), never send externally without approval (boundary), every action leaves an artifact (proof). The Xero disaster happened because there was no boundary and no proof.
12The build protocol

Bots running bots — in sprints

01
ResearchMorpheus digs before anyone builds — and again every night.
02
ChallengeChief argues the spec. Hard.
03
AlignThe human sets the North Star.
04
SpecNiobe writes pass / fail up front.
05
BuildBob builds inside the lines. Sprint by sprint.
06
ProveQA judges. Screenshot or it didn't happen.

No lost context. No rogue builds. And the point isn't many bots — it's role separation: the builder never grades his own work.

Act II — StructureSeparation creates trust
SpeakerName the crew — Morpheus, Chief, Niobe, Bob, QA. Characters beat architecture diagrams: nobody ever asked "workflow-7b, why did you double-enter Xero?" but "Bob, what happened?" starts the right conversation. Two management rules to drop here: the 5-MINUTE RULE — any agent that can't make measurable progress in 5 minutes must stop and report what it tried, what's blocking, what it needs; no silent spinning. And atomic steps — every task broken into verifiable sub-5-minute pieces. Punchline: you manage AI exactly like you manage people — most of this room doesn't do it with either. Also mention the nightly Morpheus run: while I sleep, the system reads what changed in AI and proposes what we adopt.
13Portable judgement

Rules don't generalise. Questions do.

Teach an agent what to ask and it handles work you never wrote a rule for.

01What is the owner actually trying to achieve?
02What is the customer trying to achieve?
03What tools and access do I already have?
04What proof would make this true?
Act II — JudgementObedient intern vs trusted operator
SpeakerThe four questions are the whole training program. This is what turned "minions with tools" into a team you can hand ambiguous work to.
14The system

Where everything lives — and how work moves

Work arrives The crew The workbench The gate The world
🎙️Voicenotes on the run
📧Email4 mailboxes
📄Documentsinvoices · quotes
💬Chatone thread, whole co.
Chiefchief of staff · router
🤖Minionsdomain workers
💊Morpheusresearch · nightly
🚀Niobe + Bobspec · build
🔮Oraclefinance sentinel
✅QAproof, not vibes
CRMTasksService OpsQuotesPipelineFinanceEmail hubMemory
Human gate
🌏Clients300+ managed
🧾Xero1-click finance
🚚Fieldruns · swaps · orders
Touch anything — follow the work. The dotted loop is the system learning from every day.
Act III — The workbenchBuilt, not rented
SpeakerLet it run for five seconds before speaking — the pulses do the talking. Then trace ONE journey out loud: voice note → Chief → task + CRM + draft → the gate → the client. Point at the gate: "that red bar is why I still have a business." The dotted loop underneath: retros and memory — the system gets smarter every day.
15The engine room

Models are staff, not settings

You don't hire one person for every job. Same with brains — and every lane has a fallback.

Frontier reasoningBuilds, architecture, judgement calls. The expensive brain — used where being wrong costs days.
Workhorse modelsSentinels, audits, daily operations. Fast, capable, on subscription — not per-token.
Mini modelsHeartbeats, triage, background jobs that run hundreds of times a day. Cheap and relentless.
The fallback ruleEvery lane has a plan B. The morning our main provider went dark, the whole company changed brains in under an hour — and kept working.
Act III — Vendor independenceBuild, don't rent — applies to brains too
SpeakerTwo beats: (1) matching model to job is a staffing decision — frontier brains for builds, workhorses for ops, minis for the boring constant stuff, subscriptions before API tokens. (2) The fallback story is the punchline: renting one vendor's AI means their outage is your outage. Because we built the layer, swapping brains is config, not a crisis. This is Build-Don't-Rent's strongest proof.
16The scoreboard

What it
gives back

Every line is a job that used to eat a human's week. All of it runs behind approval gates.

voice → emailai draftsinvoice automation1-click onboarding1-button ordering
Email — triage, drafts, allocation
5 h / wk
Service Ops — ordering, one button
3 h / wk
Xero — expenses & reconciliation
2 h / wk
Onboarding — quote → ops → Xero
2 h / wk
Quoting
1 h / wk
Google Ads
0.5 h / wk
+ the admin role we didn't refill
$25k / yr
Act III — Receipts13.5 hours a week, every week
SpeakerRead two or three lines, not all of them. The kicker: "That's 13 and a half hours a week — plus the $25K role we never had to refill. In a two-human company, that's not efficiency. That's survival." Then the inversion beat: hours saved is table stakes — INTERRUPTIONS REMOVED is the luxury. The system has quiet hours and one law: if I don't need to DO anything with a message, it doesn't get sent. "My AI team knows not to text me at 11 PM. Does yours?"
17What the failures taught us
›Evidence beats confidence
›Onboarding beats access
›Approval beats autonomy
›Scaffolding beats speed
Act III — CompressionEach one cost real money to learn
SpeakerRapid-fire, ten seconds each — and tie each to its scar: evidence (the app that lost a week), onboarding (the Xero disaster), approval (every gate we now have), scaffolding (the 5-week rebuild).
18Enough slides

Let's watch
it work.

15 minutes. Live system. Real business. No staged data.
Voice note in → work out. Then the field-service floor.

Demo — voice in, work outIf it breaks, that's part of the talk
SpeakerDemo 1: voice note → task + CRM + drafted email, live. Demo 2: Service Ops — client map, pass/fail visits, one-button ordering. Backup recording ready if wifi dies — and if it breaks live, own it: "THIS is why approval gates exist."
19What happens next

Follow
the build

We're building this in the open — wins, failures, cleanup days, all of it. Watch what a real business does with agents, not what a demo says it could do.

Craig · The AI Operator

Builders' group
opening soon

linkedin · /craighelmers

Act III — The launchFive minutes after this talk: follow
SpeakerOne CTA only: follow. "Connect on LinkedIn right now — and if you want inside the build as it happens, there's a builders' group opening; the link's on my profile." Don't stack CTAs.
The takeaway

Don't build
a chatbot.

Build the workbench
it can safely use.

Craig Helmers · The AI OperatorBuild, don't rent
SpeakerSay it slowly. Stop. If a plant guy can build this, the room has no excuse — let THEM think that; don't say it. This line is the ending. No thank-you slide.
1 / 18