How Stevie Bennett runs a company with a team of AI agents
The operating rules behind Helmworks, where AI agents do most of the work: written jobs, independent checks, decisions that stay with a person, and results read back before anyone calls them done.
Stevie Bennett is a software architect who has been writing software since 1992. She founded Helmworks, and she designed and built Relay, the system it runs on. Helmworks runs a team of AI workers the way you'd run a company: every job is written down, checked before it ships, and a person makes the decisions.
Here's how she describes the moment it started:
"AI did not turn me into a productivity monster. I was already that. What AI did was give the monster staff."
This page sets out the rules that make it work, with the essays where she tells each part of the story.
The short version
| Rule | What it means in practice | Read more |
|---|---|---|
| A written job, not a prompt | Every worker has a job description: what it owns, what finished work must contain, and what it may never do | I Think My AI Company Needs HR |
| Decide and act is the default | Choices, defaults, wording, order and retries belong to the worker. It decides, acts and records why | How I Became CEO of a Synthetic Company |
| Ask the owner, not the boss | When a worker can't settle something, it asks whoever owns that thing, not the founder | How I Became CEO of a Synthetic Company |
| An independent check comes first | Work is checked by someone who didn't build it before anyone relies on it | British Leyland for Robots |
| One owner per workflow | Whoever built it is the author. The owner runs it, answers for it and hands it over in writing | |
| Some things stay with a person | Anything said in someone's name, anything that can't be undone, and anything that sets how far the system is trusted | |
| Done means read back | A result counts when it has been seen where it lands: the live page fetched, the message read, the release checked | |
| The pace is set by the person | The machines never ask to slow down. The founder's capacity sets how much work is let through | How to Stop Your AI Workforce From Melting Your Brain by 5pm |
Why the founder stops being the bottleneck
The first thing that breaks when you add agents isn't the agents. It's you. Every question a worker brings to a person costs a context switch, and there are a lot of workers. In her words:
"I had become a meat-based integration layer!!!"
The fix wasn't fewer workers or slower ones. It was deciding in writing, ahead of time, what a worker settles itself, who it asks when it can't, and the short list that ever reaches her: architectural changes, anything public, anything sent, anything destructive, schema changes, new dependencies, secrets, production settings, cost. Everything else keeps moving.
"I do not want autonomy everywhere. I want autonomy where it is boring, bounded and reversible."
The organisation, not the model
Most writing about agents is about which model to use. Stevie's view is the opposite:
"It is not a chatbot problem. It is an organisational design problem."
Models change every few months. The jobs, the checks, the owners and the handovers are what make the work repeatable, and they outlast any one model. It can also go too far the other way: she has written about the week the checks became so thorough that the whole company downed tools, in British Leyland for Robots.
What it produces
Relay turns a founder's spoken sentence into reviewed, released software, and then finds and fixes its own faults.
"A sentence spoken on a Sunday morning was reviewed, released software by the afternoon."
If you want this for your own team
Helmworks sets this up for other teams: the written jobs, where the checks go, who owns what, and what stays with a person. Talk to Stevie about yours.
Stevie also speaks about this to technology and business audiences. See speaking.