What I'm doing now.

Building Hermes Agent, running Gemma 4 as my LLM entirely on local hardware.

The experiment behind it: how much capable agent you can run with no API meter ticking, which is its own answer to the usage-based-billing future I keep writing about.

Kept sharpening vinny.dev itself, now on the Signal Ledger v3.0 editorial pass, and writing into it steadily.

Five essays since mid-May, from Code Review Your Prompts to The Subsidy and the Severance, with the site staying its own working argument about consistency.

Writing a book, inspired by the conversations my blog entries keep starting.

The essays here have turned into an ongoing back-and-forth with readers and with myself, and the book is where I'm pulling those threads into something longer and more durable than a post.

"What to Make of a Life" by Jim Collins, a reflective read from the author of Good to Great on the questions worth organizing a life around.

"All We Say: The Battle for American Identity" by Ben Rhodes, Obama's former speechwriter, telling the country's story through fifteen speeches from Franklin to Trump.

A timely look at how rhetoric draws a nation's competing self-portraits, and a fitting read for a quarter spent arguing that the frame is the bottleneck.

What a design system needs when LLMs are first-class consumers: token naming that survives a code-gen pass, component APIs that don't leak Tailwind specifics, docs that double as agent prompts.

Inkwell is the current test bed.

Agentic AI for the enterprise SDLC: separating genuine productivity from tool sprawl, and figuring out which patterns actually help teams ship.

Building multi-agent and agentic workflow tools that fan out work effectively.

When code should own the control flow instead of the model (the workflow.ts pattern), how to decompose a task so parallel agents actually converge, and when a team of agents earns its cost versus just adding a tax.

Modeling adversarial review as a tournament.

Assign an agent to aggressively search a target output for flaws, then score the rounds and award points to the winners, the bet being that competition surfaces the failures a single, polite reviewer would wave through.

The reframe from 'the bottleneck is never the stack' to 'the frame is the bottleneck'.

As agents get better at execution, the constraint moves to the human framing of the problem, and which engineering investments move the work versus just feel like progress.

The economics of AI once the subsidy ends.

Usage-based billing is exposing the true cost of inference while the layoff boomerang exposes the true cost of the people, and the salary-for-subscription swap is usually fake math.

What the Anthropic release cadence does to enterprise adoption, with Design, Cowork, Chrome, Excel, and Opus 4.8 all landing inside one planning cycle.

Where does governance even attach when the surface area changes every few weeks?

AI populism as a builder's problem (h/t Jasmine Sun's warning-shots piece): what sociopolitical alignment looks like as an actual line item on an engineering leader's plate, not a press-release responsibility someone else handles.

The future of learning in an AI-native world: when models can explain anything on demand, what does it actually mean to build expertise, and how should engineers (and kids) learn differently from here?