point of view · operating thesis
My Point of View
Not a blog post. A working thesis on how I think about technology, teams, and the work of leading engineering at scale.
last reviewed · June 2026
the thesis
It was not the stack when we moved to cloud. It was not the stack when we moved to containers. It is not the stack now that we have agents. The bottleneck is clarity of intent, encoded in durable artifacts, enforced by systems rather than willpower. The leaders and builders who win this cycle are the ones who industrialize craft and bargain honestly with the humans whose careers it reshapes.
Underneath that argument are three load-bearing beliefs. Each one is fully developed across multiple posts. Each one has a contrarian edge. Expand any of them to read the case.
the thesis · three pillars
One argument, said in many rooms
01The spec is the product.
Code is becoming regeneratable exhaust. The durable artifact is the specification.
For most of the last thirty years, software organizations measured the wrong things. We measured velocity in commits, quality in coverage, seniority in lines shipped. That structure rested on a quiet assumption: turning intent into working software was the hard part. The typing was the bottleneck.
That assumption is breaking. With agentic coding tools, the act of turning a clear specification into working code has become genuinely cheap. What stayed hard, and got more visible by the day, is everything upstream. Knowing what to build. Knowing what good looks like. Reviewing whether the generated artifact actually meets the intent.
So spec review is becoming the new code review. Seniority shows up in clarity of thinking, not speed at the keyboard. The repository becomes a fossil record. The spec becomes the strategy.
“The code is exhaust. The spec is the product.”
“A vague feature shipped fast is a bug factory in slow motion.”
“If your team's generated code is wrong, you may not have a code problem. You may have a spec problem, a verification problem, a standards problem, or a review problem.”
“Standards enforced by automation beat standards enforced by willpower, every time.”
02From toil to judgment.
AI elevates work; it does not replace it. The scarce skills become problem framing, system design, taste, and the capacity to evaluate.
AI agents are a lever against the gravity of operational drudgery. In the same way that cloud freed us from racking servers and patching disks, agents free us from babysitting dashboards and chasing cost anomalies and reading the same triage runbook for the hundredth time. That is not a productivity story. It is a redistribution of human attention.
Most leadership conversations about AI miss this. They argue about model capability. The bottleneck is not model capability. It is the organization's capacity to evaluate whether outputs are correct. You need fewer generators and more evaluators. Augmented humans correct in real time. Automated errors compound.
There is a humanist version of the same idea. AI is not replacing us. It is reminding us what we are actually here for. The keyboard work was never the meaningful part of being an engineer or a leader. The judgment was. The taste was.
“The future of work will be defined by how effectively we use agents to reduce toil and reclaim human judgment.”
“Generative AI does not replace anything. It reduces the marginal cost of specific cognitive outputs.”
“The bottleneck is not model capability. It is the organization's capacity to evaluate whether outputs are trustworthy.”
“It's not humans versus machines. It's humans with machines, compounding value.”
03Build for after you are gone.
Principles over personalities. Defaults inherited, not chosen. Trust compounds faster than velocity.
If your strategy cannot survive a leadership transition, it was a preference, not a strategy. Someone's aesthetic. A personality expressed as architecture. The revolving door of leadership is not a threat to good strategy. It is a test of it.
This shows up in different language across the year. In Building Cloud From Zero, it was the discipline of making the right thing the default before anyone builds workarounds. In Pipeline-First Is Not a DevOps Initiative, it was trust-scaling, the move from individualized heroics to systemic delivery. In When Winning Teams Lose, it was accountability flowing upward, because if you publicly criticize your team while protecting yourself, you create the culture problem you are complaining about.
And in the populism essay, it became the coda I had been building toward all year. Builders set the language. Builders own the trust deficit. Vague reassurance is not safety. Specificity is. Workforce optimization is a cancer phrase.
“If you are building monuments to yourself, you are building tomorrow's technical debt.”
“Platforms only work when defaults are inherited, not chosen.”
“In the long run, trust compounds faster than velocity.”
“The best thing any of us can build is something that keeps getting better after we're gone.”
The bottleneck is never the stack.
It was not the stack when we moved to cloud. It was not the stack when we moved to containers. It is not the stack now that we have agents. The bottleneck is the frame: the constraints, invariants, acceptance criteria, and situational judgment that determine whether an agent run succeeds or fails. As models get better at execution, the bottleneck moves upstream to framing. The organizations that learn to build better frames will outperform the ones chasing better models.
Technology is an operating model problem before it is a tooling problem.
Tools matter, but they rarely fix unclear ownership, weak incentives, or disconnected priorities. Before we argue about platforms, we should agree on who owns the outcome, how work flows between teams, and what the business is actually trying to do. The same logic governs an AI reorg: redesign the workflows, decision rights, ownership, and evaluation loops before you redraw the org chart, because the org chart is a downstream artifact, not a strategy. The tooling conversation becomes easier, and cheaper, once the work is clear.
AI will reward organizations that learn faster, not just ship faster.
The advantage will not go to the organization with the most AI pilots or the biggest model budget. It will go to the one that builds feedback loops, challenges its own assumptions, and adapts the operating model as the work changes. Shipping velocity without learning velocity is just expensive motion.
The spec is the product.
Prompt engineering taught us how to talk to a model. Spec engineering is what we are learning now: a clear, versioned, reviewable specification, a layer of project standards, and a coordinated set of specialists working inside the guardrails. In an agentic SDLC, code becomes the regeneratable output of that system. The artifact your organization is producing is no longer the codebase. It is the specification. The repo starts to feel less like inventory and more like a fossil record.
Platform engineering should reduce cognitive load.
A platform is not successful because it exists. It is successful when teams can move faster with less friction and more confidence. Every platform decision should be evaluated by whether it makes the daily work of engineers easier, not by how elegant it looks on an architecture diagram. Discipline does not scale. Automation does. Standards belong in files, not in heads.
Developer experience is a business capability.
Developer experience is not an internal nicety or a recruiting perk. It shows up in delivery speed, quality, reliability, cost, and the ability to retain the engineers you want to keep. Treat it like a product with measurable outcomes, not a side project that gets attention when someone complains loudly enough.
Trust compounds faster than velocity.
Teams move faster when they trust the system, their leadership, and each other. Governance, observability, and human-in-the-loop are not friction. They are the engineering work that makes speed sustainable. When trust is present, small process problems get solved quietly. When trust is absent, even good processes produce friction. Leaders spend too much time on process and not enough on the conditions that make process unnecessary.
Leaders need to stay close to the craft.
Leaders do not need to write every line of code, review every PR, or sit in every design session. But they do need to understand how the work is changing, especially when AI, platform shifts, or new delivery models are rewriting the ground rules. There is a unique kind of empathy that comes from debugging your own code at 11 PM. Leadership three layers removed from the craft makes slower, weaker decisions.
Build for after you are gone.
Strategy that cannot survive a leadership transition was a preference, not a strategy. Platforms only work when the defaults are inherited, not chosen. If you are building monuments to yourself, you are building tomorrow’s technical debt. The best thing any of us can build is something that keeps getting better after we leave the room. That is not a small ambition. It is the whole job.
Builders owe a bargain.
Technical alignment without sociopolitical alignment is not safety. It is just better engineering for a worse outcome. Vague reassurance is not safety. Specificity is. When we describe an engineering vision in the same words we would use with someone whose career is being reshaped by it, we sound like the villain in a movie. Builders set the language. Builders own the trust deficit. That bargain is economic as well as rhetorical: the salary-for-subscription swap is usually fake math, the layoff boomerang is already rehiring the roles it cut, and the honest move is to treat AI displacement like a production change and keep your best operators close. Pick the version of the bargain you can live with.
The model cost is not the AI cost.
Enterprise AI is entering its accounting phase. The model invoice is one line item; the honest ledger also carries context preparation, review, rework, escaped defects, evaluation, governance, and institutional memory. The number worth organizing a team around is not tokens consumed or output generated, but value per verified outcome: what survived review against something more durable than vibes. A fuzzy spec used to cost time; now it costs time plus tokens plus review plus rework plus trust, because the meter did not create that cost, it made it visible. Itemize your own usage and the denominator sharpens: value per verified outcome, cost per accepted result, and portability when the preferred model disappears — with token cost demoted to a diagnostic and the cache-read ratio read as an architecture signal.
Trust the gate, not the actor.
Agents made generation abundant; verification capacity did not move with it, so review becomes the constraint. Piling on more human attention does not scale; it just recreates the four-minute code review at higher volume. The durable move is to pull review judgment out of people's heads and encode it as an enforced gate that runs every time: versioned, separate from the generator, with overrides that require a written rationale. A gate is not proof the work is correct, only proof that a standard you chose actually ran — which frees your best people to author the quality system instead of inspecting every artifact by hand.
Complexity is inevitable. Drag is optional.
Some complexity is the honest price of operating at scale: resilience, security, auditability, cost discipline. That kind cannot be wished away, and a simple system that ignores how the business actually works just hides the complexity somewhere less visible, usually onto the people least able to absorb it. The drag is the optional part — the self-inflicted weight of bespoke patterns, late-stage security reviews, costs that surface only on the monthly bill, and delivery that depends on the few people who remember where the bodies are buried. Leverage is what you design when you turn that repeated friction into platforms, paved roads, and defaults that encode hard-won lessons, so the next team inherits the judgment instead of paying for it again; a platform at its best is not bureaucracy, but organizational memory and empathy, encoded. AI raises the stakes here rather than lowering them: it changes the economics of creation, but it does not repeal the need for trust.
the arc · twelve months
Twelve months of saying it better
The thesis was earned post by post. Here is the through-line, with the line from each post that pushed the argument forward.
- May 17, 2025Back to Bloggingstart
“Welcome back, old friend.”
- Aug 10, 2025From Vibe Mode to Spec Mode
“Structure, clarity, and repeatability are what really stand the test of time.”
- Aug 19, 2025We Are Standing at the Edge of Another Tech Revolution
“Will you be ready to step into the roles that don't exist yet?”
- Sep 14, 2025From AI to AI Agentsthesis
“The future of work will be defined by how effectively we use agents to reduce toil and reclaim human judgment.”
- Sep 27, 2025Coaching Your LLM with Software Craftsmanship
“The quality of the output depends heavily on the quality of the guardrails you provide.”
- Sep 30, 2025Beyond the Hype: GenAI Adoption
“Your skepticism might be wisdom, not resistance.”
- Oct 20, 2025The Art of Enough
“Minimalism is not about doing less. It's about using what already exists more intelligently.”
- Nov 13, 2025When Winning Teams Lose: Ferrari
“Celebrate your team's wins. Own the losses yourself.”
- Dec 20, 2025The Year Ahead in AI
“In the long run, trust compounds faster than velocity.”
- Jan 3, 2026Building StillView 3.0 with Opus 4.5
“AI did not replace my judgment. It demanded more of it.”
- Mar 15, 2026Yesterday's Feature Is Today's Bug
“If you are building monuments to yourself, you are building tomorrow’s technical debt.”
- Mar 22, 2026Strip the Assumptions
“You need fewer generators and more evaluators.”
- Mar 26, 2026Pipeline-First Is Not a DevOps Initiative
“The pipeline is the artifact. The culture is the product.”
- Apr 5, 2026The Dark Factory Model
“What 30 years of experience actually gives you is a library of failure modes.”
- Apr 30, 2026Spec, Standards, Specialists
“The bottleneck is no longer the typing. It is the thinking.”
- May 8, 2026Building Cloud From Zero
“Platforms only work when defaults are inherited, not chosen.”
- May 10, 2026AI Populism is a Builder's Problemcoda
“Technical alignment without sociopolitical alignment is not safety. It is just better engineering for a worse outcome.”
- May 12, 2026The Spec Is the Productlanding
“The code is exhaust. The spec is the product.”
- May 17, 2026Code Review Your Prompts
“Engineering rigor is not a domain. It is a disposition.”
- May 24, 2026The Frame Is the Bottleneckthesis
“Models climb frames. Leaders build them.”
- May 26, 2026Workflow.ts: Code Runs the Agentlanding
“The model still did the judgment work. The code decided what counted as done.”
- May 28, 2026The Agent Team Tax
“Agent teams are not faster subagents. They are a more expensive coordination pattern.”
- Jun 2, 2026The Subsidy and the Severance
“We made implementation cheaper, so the value moved upstream to specification.”
- Jun 6, 2026Your First AI Reorg Should Be the Work
“Your first AI reorg should redesign the work, not the people.”
- Jun 8, 2026The AI Operating Ledger
“The model cost is not the AI cost. The prompt is not the workflow. The output is not the outcome.”
- Jun 11, 2026The Agent Era Gets an Invoice
“The agent became the interface in the same two weeks the invoice arrived.”
- Jun 13, 2026The Kill Switch Was Always There
“Portability moves the request. Verification proves the replacement behaves.”
- Jun 14, 2026Trust the Gate, Not the Actorlanding
“Trust the gate, not the actor.”
- Jun 15, 2026The Subsidy, Itemized
“Value per verified outcome, cost per accepted result, and portability when the stop comes.”
- Jun 18, 2026Native Was the Long Way Around
“An agent respects the boundaries your tooling enforces and ignores the ones that live only in a senior engineer’s head.”
- Jun 21, 2026Complexity Is Inevitable. Drag Is Optional.landing
“Complexity is inevitable. Drag is optional. Leverage is designed.”
signature framings · vocabulary
The vocabulary the argument travels in
These phrases recur enough across the year to be considered shorthand for the POV.
- Specs beat vibes
- The bottleneck is never the stack
- Code is exhaust, the spec is the product
- From toil to judgment
- Discipline does not scale; automation does
- Agents advise; humans decide
- Trust compounds faster than velocity
- Pair programming, not autopilot
- Polish is not correctness
- Build skills, not dependencies
- Principles, not personalities
- The pipeline is the artifact, the culture is the product
- Humans with machines, compounding value
- Augmented humans correct in real time
- Pipelines fail because urgency feels like permission
- Models climb frames, leaders build them
- The frame is the bottleneck
- Framers, not finishers
- Vibes with a diff
- Value per verified outcome, not tokens consumed
- Trust the gate, not the actor
- Complexity is inevitable, drag is optional, leverage is designed
the contrarian edges · refusals
What I refuse to say
A point of view is defined by its refusals as much as its claims. These are the six framings the argument pushes back against.
I refuse the doomer frame.
"AI takes your jobs" is investor narrative, not forecast. If there are no jobs, there are no consumers. If there are no consumers, the growth story collapses.
I refuse the hype frame.
Most of what vendors label as "agents" today are not autonomous agents in any rigorous sense. Pair programming, not autopilot.
I refuse the mandate frame.
AI tools are not failing developers. Leaders are failing to create the conditions for developers to succeed with AI tools.
I refuse the maximalist frame.
Kubernetes for one service, React for a static page, multi-cloud as a hedge. Big tools feel like progress. They are usually a tax on the next engineer.
I refuse the personality frame.
If your IT strategy cannot survive a leadership transition, it was a preference, not a strategy. Someone’s aesthetic. A personality expressed as architecture.
I refuse the corporate-speak frame.
"Workforce optimization" is a cancer phrase. "We are using AI so a smaller team can do harder work and get paid better for it" is a real plan. Pick the version you can live with.