What I built from the assistant's seat
Coming home from Dublin, four years without a job, and the agents and operating systems I built along the way. Most of the agent work was at Atlas Assistants.
Carol Kariuki
Technology & Operations Operator
I remember packing my bags in Dublin to come home. I had an interview with OpenAI that year, 2022, and it did not go the way I expected. What followed was four years without a job. I am not going to dress that up.

One of the things I did with those years was build. That is when I started vibecoding: describing what I wanted to an AI, reading what came back with a decade of hand-written code behind me, and shipping it. I built for myself first, because nobody was asking me to build for them, and the system I made to run my own days is where most of what follows started. The route back into paid work, when it came, was the one I had not planned for: executive support. Inbox, calendar, travel, follow-ups. The job title said assistant.
I took the job seriously, and I'm glad I did, because that seat turned out to be the best place I have ever sat to watch how a business runs. An executive's inbox is the business with the marketing removed. Every stalled deal, every unpaid invoice, every decision nobody wants to make passes through it. If you have spent a decade writing software, you cannot sit in that seat for long without wanting to build something.
This is what I built. Most of the agent work happened at Atlas Assistants, where the founder I support also builds AI agents, and where the brief changed from "manage the inbox" to "make the inbox run itself." But the first operating system was mine.
Carol OS, the one I built first
Carol OS came before any of the client work, though not in the form it has now. The first version had no agents in it. It was the system I built to run my own days when motivation would not: a morning briefing, a list of open loops, a weekly review I actually did. The agentic version came later, after I had seen what the founder was building at Atlas. Today it gives me a daily briefing from my real calendar, an open-loops engine that runs my weekly review, a Slack bot I message from my phone to run commands, and a heartbeat that tells me when the automation has died, because an automation that fails silently is worse than no automation. It saves me more than fifteen hours a week, and those hours are what made everything below possible.
It also has one rule I am strict about, and it mattered more once I started working inside other people's systems. It holds my interpretation of the executives I support, never a copy of their data. If their repositories disappeared tomorrow, mine would still stand. That boundary is what makes the frameworks inside it mine.
The seat
The Atlas projects only make sense against the job, so the job comes before the projects.
An assistant's week is a stream of small, unrelated things with one property in common: if nobody does them, something breaks later. A reply that sits for 48 hours becomes a lost deal. A card that expires becomes a missed payment. A meeting with no agenda becomes an hour wasted for six people.
What I learned doing this by hand is that the work is not small. It is operations, run at the resolution of a single person. The same triage applies to one founder's inbox as to any operations queue: decide what matters, route it with enough context to act on, close the loop, and write down what you learned so it never costs an hour again.
I had done this kind of work before with a whole team around me. Here I had a laptop, and a founder who was open to me building things.
I do all of this from Nairobi, for a founder in the United States. His morning starts when my afternoon is half gone, and his evening lands in my night. The gap turned out to be a teacher. You cannot interrupt someone who is asleep, so you learn to leave finished work waiting for them rather than questions, and to write everything down so the handover at the edge of each day does not depend on anyone's memory. Several of the systems below exist because the time difference made them necessary before anyone asked for them.
Why the agent work happened at Atlas
Atlas Assistants matches executive assistants with founders. The founder I support runs it alongside HUM, a house manager staffing company. His view is that the assistant of the near future is a person directing a layer of AI agents, not a person doing everything by hand. He was building that layer for his own companies, and I was the assistant sitting next to the build.
That is how I ended up with commit access to the founder's own operating system, and then to the company's. These are the projects, in the order I met them.
The founder's operating system, and Ari
The founder keeps his whole world in a private repository: businesses, coaching clients, plans, decisions, a memory layer. Claude Code reads it at the start of every session. On top of it runs Ari, a chief of staff agent that produces his daily digest, keeps a Notion archive, and takes actions inside agreed limits.
My job there started as quality control. I was the person checking whether Ari's output was true, whether the digest arrived on schedule, and whether it had done anything it was not allowed to do. Then I started fixing what I found.
The approval gate lived inside a scheduled job, so anything Ari did from a chat conversation had no gate at all. I moved it so the gate held everywhere. A rule described a situation that could not happen, so I rewrote it. Ari would produce content and then wait to be asked to post it, so producing and posting became two separate, explicit steps. The scheduled digest had gone quiet, so I restored it with the prompt kept in the repository where it could not drift.
The ordinary operations work ran through the same system the whole time: recovering an overdue invoice from a payroll provider with approval-ready drafts, or narrowing a calendar hold to exactly what the founder had asked for and nothing more.
I had never worked that way before, with my operational judgment and my engineering in the same place. It is the most honest version of the assistant job I have found.
The agent does the volume. I own whether it is right.
Ava and Hope at HUM
HUM runs two agents in production and one in shadow mode. Ava is the talent operations agent: candidate screening, SMS first contact, pipeline reporting, and a Slack bot the team talks to. Hope is the account manager agent. She drafts client-facing messages into Front, the team's shared inbox, where a person reads them before anything goes out.
My work there was the unglamorous part that decides whether an agent gets used or routed around. Hope's drafts were reaching Front with the formatting stripped, so the account manager was rewriting them by hand. I made the drafts carry their line breaks and send the fields Front requires when a draft is updated. Ava's follow-up drafts were being produced but never arriving, so I fixed the path, and then let Ava open a conversation when none existed instead of failing. A booking window counted calendar days when the team books on working days, so I changed it to count bookable days.
None of these is a headline. Each one was the difference between an agent the team trusts and one they quietly work around.
Ava on the phone
In September I built the voice version of Ava: an outbound screening agent that calls house manager candidates, introduces itself, runs the knockout questions and four fit questions, scores the answers, writes the result to the recruiting system, and alerts the recruiter in Slack. If nobody answers, it leaves a voicemail and follows up by text. It runs on Bland for the calls and Railway for the service, with a page where recruiters can start calls from a browser.
The part I am proudest of is a rule, not a feature.
The first answered call exposed three defects. A bad phone line was rejecting a candidate. The agent restarted its questions partway through. It invented a candidate's location. I fixed each one the day it appeared, and the pilot went ahead with ten live candidates, the AI disclosure switched on, and a person listening.
The Second Brain, and one executive's vault
Atlas delivers something it calls the AI Workspace Setup: an Obsidian vault wired to Claude Code, with a knowledge graph of people and companies, a writing system in the executive's voice, and a nightly process called the dream cycle. Every night it checks each escalation threshold, flags stale projects, sweeps the inbox, and writes the next morning's briefing.
In June I built one of these for a real executive who runs several businesses at once. Intake, placeholders, decision authority, escalation thresholds, the automation registry. By August the engine had been proven end to end on a sanitised demo. It is scheduled for 3am on a machine that pulls tasks from the executive's assistant, processes them, and pushes the briefing back. I wrote the handoff so the assistant who runs it day to day could pick it up without the chat that built it.
I also built the demo vault: a fictional executive on the real architecture, so the system can be shown without exposing anyone's data.
What the seat taught me
- Trust is levelled, not granted. HUM's agents sit on a ladder from observer to operator. They are promoted on two weeks of reliable output and demoted on a single trust-breaking action. It is a good way to think about agents. It is also how I would want to be managed.
- The review gate is the product. Every system here ends at a person: a recruiter, an account manager, a founder reading a briefing. The agent earns more room by being right inside the gate, not by removing it.
- The small fixes are the work. Line breaks in a draft. A gate that only held in one of two places. A booking window counting the wrong kind of day. Nobody writes case studies about these, and they decide whether anything gets used.
- Write it down so it never costs an hour again. Every one of these systems has a learnings folder, and reading it is the first step of every session. That habit came from the assistant's seat, not from engineering.
I am still in the seat. The title still says assistant. The work is building operating systems and agents for the founders I support, and that is what I want to do next, for more of them.