🧭 The Framework
"The best engineers aren't the best because they work harder — they focus on high-leverage activities when it's time to scale, and they slow down on the details when they matter most." The Agentic Engineering Operating Level is a framework built over 15+ years that maps every place you and your agent can operate — from the raw line of code up to the software factory. 0:00
The core question: where is your attention being applied? "It's always you plus your agent" — not you doing it by hand. The framework forces you to answer where your time, focus, and attention go with your agents.
"Move up and you gain leverage and speed. Move down and you gain control and understanding. You do not get both. That trade is the whole game."
🏗️ The Five Levels
Each level gives something and costs something. 3:02
L1
Code Primitives — lines, blocks, functions, types, classes. Maximum direct control; software is just individual lines of code.
L2
Code Structure — file names, modules, directories. How the codebase is organized; a bit more leverage, losing granularity.
L3
Data and Execution — database tables, databases, scripts, CLIs. "The database tables are the contracts you're building for the rest of the system."
L4
Delivery and Intent — the application (the "vibe coding level"), repository, plan, documentation. "If you can only look at the application and not above or below it, you're very limited."
L5
Agentic Systems — your agent, AI Developer Workflows (ADW), the software factory. Highest leverage; "pretty much lost control over everything."
The sub-levels that matter: within L3, scripts and CLIs are "how you control your system at the engineering level — reusable pathways for you and your agents." Within L4, planning is a "really powerful place to sit right now." At L5, an ADW (plan→build→test→deliver) is not a software factory — "it takes many developer workflows to make a software factory."
📈 Higher Is NOT Better
"Leverage on top of ignorance is a trap. You cannot scale a system you do not understand." 9:11
Point blank: building software factories or ADWs when you have no idea what your application does, what the database tables are, how things are organized as types — that ends you in a really bad place. "If you want to move up these high levels, you must first go down. You must understand the system."
But once you understand the low level of your domain, you definitely want to move up — "why keep doing work your agents can do on your behalf?" What you're looking for is a range, not a position.
🎯 Range Beats Position
The difference between a vibe coder and an agentic engineer isn't intelligence — it's how many levels you can actually see and operate on. 10:52
| Identity | Levels they can see |
|---|---|
| Vibe coder | Application only (the running app is the only artifact to inspect) |
| Knowledge worker | + documentation, plan, product |
| Data analyst | + tables, a couple scripts |
| Software engineer | + all code primitives + structure, and moving up |
| Senior / principal | + product, plan, documentation — "think through the lens of the product and the users" |
| Agentic engineer | + agentic systems, ADWs, software factory |
The career jump: Dan remembers his big leap (~2018–19) — "I started planning all my work, documenting everything, and thinking through the lens of the product and the users." That's where you go from mid to senior/principal.
⚡ Leverage & Speed (Moving Up)
Moving up gives leverage and speed — but it's the same trade as being promoted to a lead/manager/CTO: more impact, less direct control. 13:20
The cost: "we are losing understanding of the lower levels and have less control than ever over what's going on." All technology gives and takes — the question isn't if or when, but what trade-off should you make right now.
🎛️ Control & Understanding (Moving Down)
The inverse. Moving down gains control and understanding — but loses speed. 14:44
The "you're cooked" warning: "any engineer writing lines of code by hand — you're cooked, you cannot compete." (The tiny exception: high-precision C/assembly micro-optimization for HFT — and even then an agent would multiply you.) But the point is you + your agent moving down together, not you going solo.
What you lose moving down: speed and leverage — you have to pay more attention, more time. "You really want to be careful when you move up and down these scales."
📍 Where Most Engineers Are (The Two Lazy Extremes)
Most engineers oscillate between two lazy extremes: the floor (vibe coding) and the ceiling (agentic engineering). 16:00
Dan's recommendation — the levels worth more time: (1) documentation — media-rich now (images, SVGs, concise summaries, audio, short videos) so you and your agents can quickly understand what's done; (2) the database table level — "your product is your relationship with your users, and that takes the shape of data. Own your database tables"; (3) at least a baseline understanding of directory structure, files, types, and functions — going to lines is rare.
"The code is the law, the source of truth — but it's slow, and agents can do a lot of the low-level work. So we document, summarize, compress, and tell ourselves in images, audio, and short videos."
⚖️ When to Choose Leverage vs Control
The decision framework. Everything is a range, and it depends on domain too. 20:24
| ✅ Choose LEVERAGE (move up) when… | 🎛️ Choose CONTROL (move down) when… |
|---|---|
| You understand the domain (know right from wrong) | Domain expertise is thin (can't prove it's right) |
| Work is familiar and repeated — "three makes a pattern" | The system is unfamiliar (new company/codebase/project) |
| Evidence supports automation | Risk and impact are high (rockets, housing, biomed) |
| You have real agentic engineering skill | Debugging evidence is weak (no clear pass/fail signal) |
| — | Performance and taste matter (UIs all look the same now) |
| — | Work is out-of-distribution for the model |
The automation ladder: "three makes a pattern — three should be your red blinking light." Start with a skill → build a reusable agent → move it to an ADW once you want to save money and scale it.
The taste test: "all UIs are starting to look the same because everyone's vibe-slopping them." When detail/taste matters, go in — there's a signature (load-bearing, dashes everywhere) that normal people can spot.
🚫 Out-of-Distribution (What the Model Fights You On)
Out-of-distribution isn't just what the model doesn't know — it's also the behavior it actively fights you on. 25:59
In-distribution: "pretty much every workhorse-level model can build you an API, design a Postgres database, build a crappy UI." Out-of-distribution: stuff the model can't do, doesn't know, or is trained not to do.
Your expertise expands the distribution. When the model is out of its depth, you step in and teach it — context engineering/in-context learning, fine-tuning, fusion harness (multiple models, multiple plans, agents collaborating to fill gaps). "Don't collaborate with someone who doesn't know what you're trying to do — if you must, you have to teach them."
🧱 Building Expertise
"If you want to get good at something, you got to do it all the way from the atoms." The atoms of software engineering are lines, blocks, functions, types, classes, files, modules. 28:02
"There is no replacement for raw expertise. There just isn't. It's raw time — bang your head against the computer. Then, and only then, can you point yourself and your agents at the high levels and steer them at extreme leverage."
Agentic engineering is pointless if you don't understand what you're agentic-engineering at some level. The depth you need depends on: your expertise level, how risky it is, and how much your agent already knows.
🚀 Beyond the Factory (Dark Factory + RSI)
Levels above the software factory. 29:53
| Level | What it is |
|---|---|
| Agent | Industry standard now — "just using an agent gets you average results" |
| ADW (AI Developer Workflow) | Repeatable workflow multiplying the SDLC with agents — "not the loop, the workflow" |
| Software Factory | Composed ADWs — not just SDLC, but hot fixes, feature prototypes, ideation, staging launches, support tickets, "the system that starts working without you" |
| Dark Factory | Next tangible level — living software that runs while you sleep |
| RSI | Non-existent state-of-the-art — "the leaders (OpenAI, Anthropic) are trying, no one's doing it yet" |
"Nobody jumps straight there." You don't jump to RSI, the dark factory, or even the software factory. You start at the low levels, understand, then scale and combine. "You earn it by understanding every level underneath."
The announcement: Dan is building the Phase 3 successor to Tactical Agentic Coding (Phase 2 targets the top 20% of engineers, "build the system that builds the system"). "Our mission: build living software that runs while we sleep."
📚 References: What Is Agentic Engineering? · Understand the Software Factory · Forget Loop Engineering · Fusion Harness Engineering
💡 Key Takeaways
- Higher is NOT better. Leverage on top of ignorance is a trap — you cannot scale a system you don't understand.
- Move up for leverage and speed, down for control and understanding. You don't get both — the trade is the whole game.
- Range beats position. The difference between a vibe coder and an agentic engineer is how many levels you can see and operate on.
- "Three makes a pattern." Your red blinking light to automate — skill → agent → ADW → factory.
- The data level matters. Your product is your relationship with users, and that relationship is data. Own your database tables.
- Out-of-distribution isn't just what the model doesn't know — it's what it actively fights you on. Your expertise expands the distribution.
- You earn the high levels. Nobody jumps to the dark factory or RSI — start low, understand, then scale.
📍 Timestamp Index
0:00 Time, focus, attention
1:04 The framework
2:57 L1 — Code Primitives
3:57 L2 — Code Structure
4:30 L3 — Data & Execution
5:38 L4 — Delivery & Intent
7:43 L5 — Agentic Systems
9:11 Higher is not better
13:20 Leverage & speed
14:44 Control & understanding
20:24 When to choose leverage vs control
25:59 Out-of-distribution
28:02 Building expertise
29:53 Dark factory & RSI
34:47 Your operating level