Most people install an AI agent, use it as a chatbot, and conclude it saves them a few minutes a day. They are operating a system with 15 levels at level 1. This piece is the full map: what each level does, what it unlocks for your business, how to set it up, and the mistake that kills it, all the way from a first prompt to a system that watches your business around the clock and only bills you when something actually happens.
The levels run on Hermes Agent from Nous Research, an open agent framework you install on your own machine. My company's agent, a 38Shift fork of Hermes, runs in production, so where I have receipts I show them. Everything here also works with any agent that can run commands and manage files.
Phase 1: Foundation (levels 1 to 3)
Level 1 is the shift from searching to assigning. An agent is not a chatbot: it works across your files, your terminal, and the web, which means it can deliver finished work instead of answers you still have to act on. The mistake that keeps people stuck here is treating it like a search engine: "tell me about the best CRMs" gets you a wall of text to read and act on yourself. There is no setup at this level. Paste this instead and you get a deliverable:
Research the top 5 CRMs for solo founders, compare pricing
and features, save the report to a file.Same tool, different sentence, and the second sentence is the whole point.
Level 2 is SOUL.md, one file that changes every answer the agent ever gives. It is a plain text file the agent reads on every single turn: who it is, how your business works, what it must never do. Fifty to eighty lines covering identity, voice, operations, and restrictions is enough. The difference is not cosmetic. Ask "should I raise prices?" with an empty SOUL.md and you get generic pricing advice. Ask with a filled one and you get an answer computed against your actual margins, your actual customers, and your actual positioning, because the agent knows them.
My production agent runs on a 51-line SOUL.md. Two of its rules, verbatim: "Never guess missing data" and "A rejected operation is better than a wrong one." Those two lines are why I can let it touch real business systems. Here is a starter frame you can paste and fill in:
# SOUL.md
## Identity
[Agent name], operations agent for [your company].
## Business
What you sell, to whom, at what price, current focus.
## Voice
Concise. No filler. Report results, not process.
## Rules
Never guess missing data.
Confirm before anything leaves the company.
When uncertain, escalate with the specific issue.The mistake: leaving it empty. An agent with no SOUL.md is a generic assistant that happens to have your login.
Level 3 is the four commands almost nobody uses, and they are the difference between running your agent like a single-lane road and running it like a highway. /background fires a task without blocking the conversation, so you can keep drafting a proposal while the agent researches a competitor, and five minutes later the analysis appears without you ever breaking flow. /steer redirects the agent mid-run without killing the work in progress. /queue lines up the next task to start the moment the current one finishes. /model switches brains mid-session, so you plan on a premium model and execute on a cheap one. The mistake is simply not knowing these exist: most people type a prompt, wait, type another prompt, and wonder why the agent feels slow.
Subscribe to Premium to read the rest.
Become a paying subscriber of Signal to get access to this post and other premum subscriber-only content.
Upgrade