Kai does not sleep, does not lose track of a thread, and checks his own work before anyone asks him to. He is an AI agent with a name, a personality file, and a narrowly defined job inside a growing stack of hundreds like him. By the creator’s own count, that stack now handles 92% of all the work running across his companies, a figure that took shape not through some overnight automation blitz but through a methodical, stage-by-stage process that anyone with a clear outcome in mind can replicate today.
Chat versus agent, and why the difference changes everything
Most people using AI right now are using it like a very fast search engine. They type a question, they get an answer, they copy the answer somewhere useful, and then they do it again tomorrow. That is chat: the tool waits, the human prompts, the tool responds. An agent flips the dynamic. It pushes rather than pulls, running workflows on a schedule, checking its own output, and looping back to improve without waiting to be asked.
The creator breaks the internal logic of an agent into four functions he calls DATA. An agent can diagnose a problem and solve it on behalf of the user, the way a consultant would. It can assemble a plan and identify the tools needed to execute it, functioning more like an architect. It can take action, executing the actual task. And it can assess its own output, reviewing whether the result landed correctly and correcting course if it did not. That last step, the self-assessment loop, is what separates an agent from a simple automation. Without the loop, a tool does the job once and stops. With it, the agent keeps getting sharper.
Before building anything, there is a quick filter worth running called the rule of R: is the task repetitive, does it follow a rules-based process, and does it generate a meaningful return on the time invested to build it? A two-minute task that would take two weeks to automate stays manual. Everything else is a candidate.
The AGENT framework for building one step at a time
The build process itself follows a five-step acronym. The A stands for aim at a specific outcome. Rather than defining every step the agent should take, the instruction is to define the destination and let the agent find its own route. A definition of done for an inbox agent, for example, might read: every morning at 9:00 a.m. the inbox is empty, replies are drafted in the user’s voice, anything that needs a human decision is flagged to the top, and nothing important slips. If the outcome cannot be stated in one sentence, the build is not ready to start.
The G is for give it an identity, delivered through three plain-English files. A soul file defines personality and tone, a no-corporate-fluff directive, a preference for flagging rather than guessing, a ban on the phrase ‘I hope this email finds you well.’ An identity file gives the agent a name and a tightly scoped role. The example built here goes by Amelia, or Emailia, as the creator put it, which is the kind of specificity that actually makes the agent more effective. A published study he cited found that stripping identity files from an airline customer support agent dropped its success rate from 33% to 11%, the same model, the same task, three times worse because it had lost its sense of who it was. A user file rounds out the trio, giving the agent context about the person it works for, their priorities, their companies, their communication habits.
The E is equip it with context and tools, and context is described as the moat. The recommended approach is to let the agent reverse-engineer its own playbook from existing source material. For an inbox agent, that means connecting it to Gmail, reading 50 sent messages, studying tone, greeting style, sentence length, and signature habits, then having it draft replies to unread emails to test what it learned before any of it is locked in.
The N stands for narrow the scope, which is where most first attempts break down. Loading a single agent with 17 tasks is the fastest way to create what the creator calls context rot, a cluttered desk where nothing gets handled cleanly. The working model here uses a manager agent, one whose only job is coordinating sub-agents, each of which handles exactly one specialized function. Kai, the orchestration agent, manages a research agent, a relationship agent, a coding agent, and a reporting agent, then pulls the results together and delivers a single answer upward. One agent to talk to; everything else handled below.
Model selection matters at this layer too. Simpler, high-volume tasks like sorting and labeling run on lighter models. The inbox agent runs on Sonnet because, as the creator noted, ‘I don’t need an Opus-level genius to run a process that we’ve already defined.’ A code refactor that might have cost 150 dollars on a more powerful model ran on the cheapest tier for a dollar fifty.
Amelia’s inbox, and the moment the leash went slack
The final step is T, for trust, built in stages. The first pass is read-only, letting the agent sort. The second pass is draft-only, reviewing every reply before it goes anywhere. The third loosens the send permissions on a narrow category, forwarding finance emails, routing Slack notifications to a label. The fourth sets a recurring heartbeat, every 15 minutes instead of once a day, because once the drafts are reliable, waiting until the next morning is just friction.
When the creator walked his executive assistant through the finished inbox agent, she assumed her job had just been handed to a machine. Instead, the sorting, the drafting, the triage summary that used to fill her mornings moved to the agent, and she moved to higher-level project work. The system then rolled out to the full team.
Kai at the desk, three files on the surface
Somewhere in this system, an agent named Kai still has his soul file, his identity file, and his user file arranged the way a person arranges a desk: the things used most on top, the memory filed below where it does not create clutter, the tools to the side ready to connect outward. He picked his own name, offered his reasons for it, and then updated his identity file to make it official.
A study cited at the outset of this journey projects 170 million new jobs created by AI by 2030, most of them not chatting with the technology but building and directing it. Kai, sorting emails every 15 minutes without being asked, is already doing the work.
