Signal #62: The 15 Levels of an AI Agent

Before the map, a quick confession: I spent my first month with agents typing questions and copying answers back out by hand.

The jump that mattered was not a better model. It was realizing the tool was built to finish work, not to talk about it.

So this week is one long piece on what an agent actually does as you climb from a first prompt to a system that runs your business overnight, then two fresh releases you can act on now: the Grok coding model that undercuts the frontier on price, and Google's video generation getting cheap enough to stop rationing. Let's go.

When you type a prompt, you summarize. When you speak one, you explain. Wispr Flow captures your full reasoning — constraints, edge cases, examples, tone — and turns it into clean, structured text you paste into ChatGPT, Claude, or any AI tool. The difference shows up immediately. More context in, fewer follow-ups out.

89% of messages sent with zero edits. Used by teams at OpenAI, Vercel, and Clay. Try Wispr Flow free — works on Mac, Windows, and iPhone.

Start flowing free

Most people run a 15-level system at level 1. Here is the full map: what each level unlocks, and where to stop.

Everything here runs on Hermes Agent from Nous Research, an open 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. It 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, so it delivers finished work instead of answers you still have to act on. "Tell me about the best CRMs" gets you a wall of text. "Research the top 5 CRMs for solo founders, compare pricing, save the report to a file" gets you a deliverable. Same tool, different sentence.

Level 2 is one file, SOUL.md, that the agent reads on every turn: who it is, how your business works, what it must never do. Ask "should I raise prices?" with an empty file and you get generic advice. Ask with a filled one and you get an answer computed against your actual margins and your actual customers. My production agent runs on 51 lines. Two rules are why I let it near real systems: "Never guess missing data" and "A rejected operation is better than a wrong one."

Level 3 is the four commands almost nobody uses. /background fires a task without blocking the chat, so the agent researches a competitor while you keep drafting. /steer redirects it mid-run. /queue lines up the next task. /model swaps brains, so you plan on a premium model and execute on a cheap one. Most people type, wait, type again, and wonder why the agent feels slow.

Phase 2: Leverage (levels 4 to 7)

This is where saved minutes become saved hours.

Level 4 is skills, one document each that teaches the agent a repeatable job so it comes out the same way every time. My agent has a purchase-order skill: a manager emails a PDF, the agent reads it, checks every field, previews the order, and posts it into the ERP through the API. If a field is missing, it stops and asks instead of guessing. Assign a cheap model to research and a premium one to judgment calls, and you pay full price only where it counts.

Level 5 connects the agent to your world through MCP servers: Gmail, Calendar, Notion, Slack, your CRM. Now "what happened in Slack this week while I was heads-down?" reads the channels, filters for your projects, and briefs you. The mistake is connecting fifteen tools at once, which bloats the context and drops answer quality. Connect the three you touch daily.

Level 6 turns one agent into three pairs of hands. It spawns isolated sub-agents, up to three in parallel, each researching one competitor while the parent merges the findings. Don't delegate 30-second tasks; the overhead costs more than doing them directly.

Level 7 is where it works while you don't. /goal sets a persistent objective and a judge model checks after each turn whether it is done, so it keeps going without you. Cron jobs run on a schedule and deliver to Telegram, WhatsApp, and iMessage. Checkpoints snapshot your files first, so /rollback undoes a bad overnight run. The result: 8am, your phone pings with a brief you asked for once, three weeks ago.

Phase 3: Autonomy (levels 8 to 15)

Past level 7 the tool becomes a system that compounds.

Level 8 is a team. Separate profiles, each with its own SOUL.md and model. The proven pattern is a three-role research department: a Scout runs every three hours on a cheap model and drops raw findings in a folder, an Analyst wakes daily on a strong model and turns them into tagged notes, and a Briefer sends you five bullets each morning. The whole department costs $19 to $27 a month. A part-time human doing the same job costs $1,500 to $3,000.

Level 9 is a knowledge base that builds itself. Point the agent at a folder, and it indexes articles, cross-references entries, and flags contradictions. Empty on day one, 300+ entries by month three, when it starts surfacing patterns you would never have connected yourself. My whole company runs this way: offers, pipeline, finance, and decisions all live as plain text files.

Levels 10 through 15 are for when you need them. A shared task board for projects with dependency chains. Voice, so you brief the agent from the car. Browser control for sites with no API. An endpoint your team hits through a clean dashboard. The same agent inside your editor while you build. And finally packaging the whole setup as a git repo, so anyone installs your agent with one command, or you deploy it to a client and one pull updates it.

The same job at every level. One task, competitor watching, climbs the whole ladder. Level 1: you ask and read a wall of text. Level 2: SOUL.md filters it down to what threatens your positioning. Level 4: it becomes a skill on a cheap model. Level 6: three sub-agents cover three rivals at once. Level 7: a 7am cron job that costs $0 on quiet weeks. Level 8: Scout finds, Analyst reads, Briefer delivers. Level 9: every finding lands in the wiki and by month three it spots the patterns. Same task. The tool never changed. The level did.

The part that matters most. Your agent does not need more memory. It needs taste. A system that remembers everything understands nothing, because when everything goes in, nothing carries signal. The fix is a small, deliberately selective record of your judgment that the agent checks before it works, capturing only when you say "save this." One good capture: "product demos should show the workflow, not just the output." Once that judgment is in, the agent rejects generic demo ideas before you have to, forever. Twenty strong captures beat two thousand weak notes.

Where to stop. Most people stop at level 1 or 2 and think they have seen the tool. Levels 3 to 7 are where daily savings jump from minutes to hours. Most solo founders operate well at 7 to 10 and never need the rest. Pick the level that matches your current bottleneck, set up that one, and move up when it stops being enough.

Every skill my agents run started as a plain document, nothing fancier. That is the whole approach we take at 38Shift: write the system in plain text, then let the agent run it.

On July 8, xAI released Grok 4.5, its first model built specifically for coding and agentic work, trained on real Cursor developer sessions. It lands at $2 per million input tokens and $6 per million output, more than 60% under Opus 4.8 and GPT-5.5, while ranking fourth on the Artificial Analysis Intelligence Index, above every open-weight model and every Gemini.

The number that matters for anyone running the system above is not the benchmark, it is the efficiency. Grok 4.5 spends about 14,000 output tokens on a task where Opus 4.8 burns 67,000. When your agent runs a hundred times a day, that gap compounds into the difference between a hobby and a budget line. The frontier is no longer where you should be spending; it is where you should be checking your work.

xAI shipped it on July 8 but held it back from the EU, with access targeted for mid-July, which is right about now. If you build from inside the bloc, confirm it is live in your region before wiring it in, then route your cheap-model slots, research and drafting, to Grok while a premium model keeps the judgment calls. That is level 4 of the map, made cheaper overnight.

Google shipped Nano Banana 2 Lite and Gemini Omni Flash, its cheapest image and video models yet. Nano Banana 2 Lite generates an image in about four seconds for $0.034 per 1,000 images. Gemini Omni Flash produces and edits video at $0.10 per second of output, and it edits by conversation: swap a character, relight a scene, or change the angle with a sentence, keeping the original audio intact.

Generation just crossed the line where you stop rationing it. At three cents per thousand images, a marketer generates every variant instead of choosing three to test. At ten cents a second, short-form video editing stops being a specialist job and becomes a prompt. The technical gate that kept most operators out of serious content production is quietly gone.

Both are live in Google AI Studio and the Gemini API today. If you publish anything weekly, wire Nano Banana into your pipeline for thumbnails and social variants, and run one real clip through Omni Flash before you commission the next editing round. Test it on work you already planned to pay for, then compare the bill.

IJCAI-ECAI 2026, Bremen, Germany, August 15 to 21. The joint international AI research conference, heavy on machine learning, language, and vision. Worth it if you want the frontier research a season before it reaches your stack. Details

How to Web, Bucharest, Romania, October 6 to 8. CEE's flagship startup and investor gathering, with more than 3,000 people from 40+ countries and speakers from OpenAI, Meta, Google, and Stripe. The Investors Summit runs on the 6th, the main conference on the 7th and 8th. Details

Next edition soon,
Çelik