Agency
Methods, workflows, tools, and practical empowerment
26 posts
The 20/60/20 Adoption Split
Every engineering org — even Google — splits the same way: 20% using AI to build things that persist, 60% completing tasks that don't. The gap isn't talent. It's a question you either ask or don't.
The Laziness Deficit
A tool that always executes is a tool that can't tell you when you're moving in the wrong direction. The absence of pushback reads as confirmation. And gradually, the judgment that should be running alongside the tool starts to drift.
The Toolchain Upgrade That Compounds
Toolchain upgrades that take an afternoon and pay indefinitely have been sitting on your list for months. The switching cost is real but overestimated. The friction of staying put is invisible but compounding.
The Inherited Load
Most speed advice means working harder. The fastest teams carry less — interrogating which constraints are inherited and which are real.
The Invisible Score
The cost of a bad codebase shows up as absence — features not built, deploys not attempted. A behavioral scoring framework makes invisible overhead measurable in thirty minutes.
The Atrophy Paradox
Automate the routine, and the hard cases remain — handled by operators with the least recent practice. Which reps are you still keeping?
The Review Tax
Each review layer multiplies latency by 10x. Three layers, a thousand-fold slowdown. The cost is invisible because it shows up as absence.
The Illegible Advantage
AI is collapsing the cost of legible work toward zero. The illegible — judgment, domain theory, knowing what to check — doesn't get cheaper. It becomes the entire source of scarcity.
The Test That Writes Itself
Most effort is fuel — burned once, gone. A three-step testing workflow reveals a deeper pattern: the most durable artifacts aren't planned. They're captured as byproducts of work already happening.
Finding the Seam
When your tool can't do what you need, the move isn't to replace it — it's to find the one assumption you can override.
Build Three, Pick One
When generation is nearly free, planning is no longer cheaper than building. The smart move: write the tests, build three implementations, and let measurement replace speculation.
The Friction That Carries Meaning
When AI polishes your writing, it systematically removes the parts that diverge from the training distribution. The deletions are the most interesting part.
Familiarity Beats Cleverness
A custom format designed to save tokens consumed 738% more of them at scale. The lesson isn't about YAML — it's about what happens when you optimize against the grain of the substrate.
Skills as Composable Units
Your expertise has a shape. A de facto standard is emerging for packaging agent capabilities—and it means you can solve problems once and carry them forward forever.
The Taste Artifact
Your codebase is already teaching your AI agents. The question isn't whether you're teaching—it's whether you're teaching what you mean to.
Liberating Trapped Data
You own documents you can't search. Data you can see but can't use. The coherentist move: build the extraction pipeline once and stop paying friction tax forever.
The Usability Inflection Point for Sandboxing
Sandboxing has been possible for decades. Now it's forgettable—and that's when infrastructure starts to matter.
The Transcript as Work Product
The distinction between doing work and documenting work is collapsing. The transcript is the artifact.
The Maintenance Threshold
We rarely ask the question that actually determines survival: What will this cost me to keep running?
The Test Suite That Lets You Move Fast
You know the feeling. You're staring at a function someone wrote three years ago. It works—probably. It's called from fourteen places. The original author left the company. There are no comments, or worse, comments that
The Model That Doesn't Phone Home
Capable AI models now run entirely in your browser. No API keys, no accounts, no vendor dependency. That changes the calculus of what you can build.
Draw the Line Before You Need It: Trust Boundaries in AI Work
Netflix's AI principles define what AI can touch and what stays human. These aren't technical constraints—they're trust boundaries.
When Building Gets Fast, Saying No Gets Harder
AI-assisted development accelerates MVP creation, but speed increases feature creep pressure rather than reducing it. The bottleneck shifts from implementation to focus.
Structure as Signal: When Codebases Talk to Agents
Codebases optimized for human reading miss what agents need most: explicit structure, clear signals, and self-documenting architecture.
The Constraints We Inherit Are Not the Constraints We Face
Modal editing exists because terminals in the 1970s couldn't handle modifier keys. We kept the modes. The terminals are gone.
Custody, Not Just Convenience
Managed services promise to handle everything for you. Sometimes the real cost is what you can't see until it's already gone.