What's Good for Humans Is Good for AI

Never spend more time arguing over who should do the work than it would take to do the work.

For the last few years my signature line in my email has been "The best measure of progress is working software." I had a few motivations for adding that to my signature, but the main one was that "show don't tell" was the fastest way to silence the naysayers and separate activity from outcomes. As I progressed further upward in my career, I also found myself further away from the code. I always prided myself on being and staying technical, but there just weren't enough hours in the day to be hands-on-keyboard and build anything of substance, even with AI. That was, until fall 2025 when vibe coding really shifted to agentic coding. I could build again, because now I had an agent that could run on the side.

So when I found myself on a call listening to a debate about who should do a coding task, I didn't jump in right away, I just kicked off an agent to work on the task. As the debate shifted to whether or not this should even be done, I refined the code with the agent and evaluated the functionality. It wasn't perfect, but it was functional, and because of my system prompts, it was well-structured code I could iterate on. I interrupted the debate by sharing my screen, demoing the prototype screens, and showing, not telling.

Over the next couple weeks it became my hobby project. I expanded it to be the front-end, back-end, database, synthetic seed data, OAuth, observability, CI/CD, accessibility, themed, the whole thing. Not my day job. My little side project, with my side agents working on it.

The original debate had been happening because of the implicit time and cost for building software, and the mental muscle memory that developed around it. Discussions, wireframes, and pixels were cheap in comparison to engineering time that was expensive, costly if done wrong, and long in duration. We're re-defining the laws of software engineering physics.

The environment is the variable

B = f(P, E). Behavior is a function of the person and their environment.
Kurt Lewin, 1936

Think about the best experiences that you've had over your career. The managers and leaders who had massive positive impacts on you. Everyone has had those memorable, impactful experiences that helped them grow and realize their potential. Where you were set up for success, and able to realize it.

Think about the environment. An environment where you were able to thrive, not just survive. Earned trust, clear expectations, flexible and supportive structure, constructive and informative feedback, empowerment and enablement. Access to tools, infrastructure, job aids and documentation. Guidance without being overbearing. Information that informed without overpowering.

Now contrast that with the bad experiences. The times where you weren't set up for success, where your engagement was low, and your performance under a microscope. Mismanaged expectations, no structure, lack of clear outcomes, unclear guidance, ambiguous or no feedback, micromanagement, low trust, learned helplessness. Missing or inadequate tools, lack of supporting infrastructure.

The person is the same. The environments are quite different. And the same things that create an environment where humans thrive, have the same constructs of an environment where AI is the most powerful, guided by and augmenting humans. That's the central thesis of this article. What's good for humans is good for AI.

The slight corollary: what humans are good at, AI is not necessarily good at. That's the hard stuff, the creative stuff, the squishy stuff, and that's what we as humans should never delegate to AI.

The reason is that AI is highly structured. Modern models are trained using tens of trillions of tokens, and hundreds of billions of parameters. It can simulate a lot, but essentially what we're doing when we prompt an LLM is constraining that vast network of parameters, zeroing in on a subset, which then gives us probabilistically better output. So the better the structure, the clearer the instructions and guidance, the better the model will lock in on our objectives.

Six dimensions, same for both

Here are the six dimensions where humans and AI need the same things, and where friction for one is friction for both.

1. Clear instructions and structure

Ambiguity is friction for humans and hallucination fuel for AI. Clear requirements, well-structured inputs, explicit expectations. When a team struggles because the requirements are vague, an agent will struggle for the same reason, and produce confident garbage instead of asking for clarification.

2. Identifying goals and outcomes

Humans need to know what "done" looks like. So do agents. Acceptance criteria, definition of done, measurable outcomes. Without them, humans gold-plate and agents over-generate. In Agile development, having a clear "Definition of Ready" and "Definition of Done" are foundational team working agreements. The litmus test is "how will I know that I'm done, and how will I test that." If you can't answer that question, regardless of the task, then you are lacking guidance and leadership.

3. Feedback and guidance

The gift that keeps compounding. Humans improve through feedback loops. Agents improve through feedback loops. The tighter the loop, the faster both improve. The agentic loop (Reason, Act, Observe) is the AI version of the same principle that makes human teams get better over time.

4. The ability to safely try things out

Psychological safety for humans, sandboxed environments for agents. If people are afraid to experiment, innovation stops. If agents can't safely test against real-world conditions, quality stops. Feature flags, staging environments and rollback capabilities help human and agent teams alike.

5. Enablement and empowerment

Give people the tools, access, and authority to do the work. Give agents the context, permissions, and contracts to operate. Gatekeeping is friction for both. The organizations that empower compound. The ones that gate-keep stall.

6. Every friction point is shared

When you find friction for humans, you've found friction for AI. When you remove it for one, you remove it for both. This is what elevates the thesis from an engineering insight to a leadership principle, and it's the through-line for everything that follows.

What AI does better than humans

AI loves toil.

It excels at finding a needle in a needle stack. It is a genius in arcane but powerful syntax like regular expressions, command line tools and low-level protocol details. Hunting through logs, traces, and stack dumps. The mountains of information that we typically have to comb through to diagnose an issue? AI is a savant at pulling that together. Let it.

AI is great at following rules. It does better with guardrails, Standard Operating Procedures, regulations and legalese. The more structured and rule-based the information, the better the output. For all of that content that's a slog of inter-relation and inter-dependency, encapsulate it as skills and prompts and delegate it.

The "lather-rinse-repeat" tasks. Those that you briefly fantasize about automating, but then just muscle through because it's "faster", AI will build you a tool to finally automate it for you.

The 100x gap: the receipts

The previous post established the arc. Three eras, the compounding equation, the urgency. This is where we show the receipts.

Going from manual to automated is 10x. That was the promise of the original Frictionless Enterprise. Going from automated to AI-native is another 10x. The math is multiplicative. These aren't projections, they're current measurements. And the number is growing daily.

If you're still trying to scale with humans, you're 100x less efficient.

The original Frictionless pyramid still holds. Same shape, radically different cost curves. The foundation (Flow, DevOps, QE, Architecture) is still the base. The domain stack is still the structure. But every layer's economics have been massively disrupted. What used to cost $100k now theoretically could cost $1k if coding was the bottleneck. The timeline for coding is collapsing from months to hours. Realizing this math is going to require a complete re-tooling of the software delivery life-cycle, because in wait states the meter is still running.

The spectrum of outcomes is real, following the adoption curve from Crossing the Chasm, skeptics at one end, compounders at the other, most organizations in the messy middle at 10-15% utilization. The frontier is already at 3-10x force multiplication. The bimodal split is widening exponentially, not linearly.

What comes next

This isn't just about coding. Software is the first domino. That was the point of the opening post. The six dimensions of "what's good for humans is good for AI" apply to product management, design, security, data, operations. Every discipline.

Everything that follows is the applied playbook. The structural reason the gap exists. How the frictionless pyramid shifts. Why complication kills every component of the rate function. How to actually move from the messy middle to compounding.