Initial experiments reveal that Open AI's subscription plan is heavily subsidized: • Subscription plan: $100/month GPT Pro • API cost: ~$550/week or ~$2,200/month

My 2-hour AI migration result: • 1,918 lines of F# became 2,359 lines of Rust • The app built and ran locally • But deploy was borked and AI added a _lot_ of verbosity to the site copy

Humans drive the agent, agent drives the tools. - Humans - Declare intent / purpose - Agents - Decide how to accomplish it, fan out to tools - Tools - Provide deterministic info / feedback Improve the system each cycle. hamy.xyz/blog/2026-06...

AI creates 100% of my code now. I haven't written or edited a line of code in months. The closest I get is writing my blog posts which I happen to do in the same editor (nvim btw).

For one-shot prompts Ruby, Python, and Go are 1.5x faster and cheaper than Rust (ai-coding-lang-bench). But real agentic work is iterative. That's where Rust's compiler loop (expressive types, ownership) shines. These guardrails scale correctness to large codebases, orgs, and teams of agents.

I'm trying to lean into a day/night cycle with agents where I let them cook while I'm AFK. I close loops at 16:00, check agents again at 20:00, then review their overnight work the next day.

My average coding throughput across 3 phases: • Before AI agents: 7 PRs/week • Early agents: 12 PRs/week • Current workflow: 20 PRs/week That's roughly 3x my manual baseline and almost 2x my early agent workflow.

I tried Rust years ago to build a simple CSV parser and got absolutely nowhere in an hour. Then rewrote the same thing in Python in 15 minutes. But fastforward to 2026 and AI is now a 24x7 tutor and coworker which vastly improves the infamous learning curve.

OpenCode web can power the same build-from-my-phone workflow as Claude Code remote control, but setup is more manual: • Secure access via Tailscale • Scripted startup + recovery It works, but I wouldn't recommend it unless you're comfortable scripting your computer.

My current AI model split: • GPT 5.6 Terra / Sol for general coding and reference work • Claude Opus for subjective writing and UI • Claude Fable almost never (too slow + expensive, only for truly hard things) • Human review for the final acceptance / translation

My first 3 days with OpenCode + GPT 5.6 Sol on Open AI's $100/month GPT Pro Plan: • 75% of weekly quota used • 407M logged tokens • ~$413 API-equivalent usage at ~90% input cached.

My early GPT 5.6 vs Claude results: • Tasks that would take Claude 30s-minutes to return data for are almost immediate • Usage looks 2-3x cheaper • Sol is practical enough to use way more often than Fable Early anecdotal data but looks very promising for a daily driver.

My AI setup is becoming model agnostic: • Terminal-based workflows • Skills shared across harnesses • Models and providers that are swappable via config

You drive, let agents do the footwork. - Humans - strategy, vision, systems, and guardrails. - Agents - research, reading, writing, and coding. Keep to 1-2 tasks at a time, use agents to fanout underneath. hamy.xyz/blog/2026-06...

The AI labs tend to one-up each other every few months. A model-agnostic setup lets me keep the same workflows and swap the model or harness when the leader changes. I think that flexibility will prove to be a competitive advantage over time.

- Go: fast, but no sum types allows more logic bugs - TypeScript: huge ecosystem, but types are unsound (Zod, any, as x) - OCaml/Haskell: great types, but AI writes non-idiomatic code - Rust: expressive types, compiler hints, and AI writes well

What I actually spend on AI as an engineer: • Work: ~$4k/month (~$50k/year), nearly all Claude Opus • Personal: ~$200/month on Max 20x. The difference is mostly hours / day (8h at work, 0-2h personal) and having access to subsidized sub pricing.

I built a simple calculator for how I'm thinking about AI token spend ROI. Example calc: - Your cost: $100k - AI cost: $50k - Speedup: 1.6x - Your value with AI = $100k x 1.6 = $160k - AI value = $160k - $100k = $60k - Net value = $60k - $50k = $10k hamy.xyz/blog/2026-07...

SWE-bench Multilingual tested agents across 9 languages and Rust came out on top for task completion. Rust isn't the fastest or cheapest for AI generation - Python, Go, and Ruby win that. But the extra time spent iterating against types/compiler/linter turns out to be a feature, not a bug.

AI is getting so good at writing standard code that it starts to look like a natural language abstraction on top of programming languages. This changes the calculus from why would I use x to why would I not use x? Rust has great perf, types, and ecosystem. With AI, it has great devx, too.

TL;DR - In 2026.H1, I started feeling like a real dad, took parental leave, went to a programming retreat, shipped CloudSeed Rust + QuietCut, reached 130k views, got back to building for fun, and started an AI diet. hamy.xyz/blog/2026-h1...

Rust's learning curve is overexaggerated. Most of the pain comes from fighting the borrow checker when you don't need to. Write high-level Rust — immutable data, cheap clones, service traits — and it all fits together.

My favorite way to organize hypermedia endpoints is Backend for Frontend: - /PAGE - renders full page - /PAGE/ACTION - Performs action, renders appropriate component Reuse coarse-grained components across page actions. Rust + Datastar - hamy.xyz/blog/2026-03...

Getting Started with Datastar - Build a Rust + Axum Todo App - HAMY

Essay - Published: 2026.04.08 | 8 min read (2,048 words)build | create | datastar | rust

hamy.xyz

Rust is a great high-level language if you abide by a few rules: - Immutable data - Functional pipelines - Cheap clones - Arc dyn traits for dependency injection After a couple weeks (and the help of AI) I got the devx competitive with C# and TS.

I think talking will become the primary means of driving computers simply because it's the path of least resistance. I think of my mom, brother, and wife who still poke type and how much easier it's going to be for them to talk vs type large text blobs.

Anecdotal results of agentic engineering: - Before: I shipped 1-2 PRs a day, averaging ~7 a week - After: I ship 2-3 PRs a day, averaging ~12 a week So my guesstimate is a 50-60% speed improvement over my previous workflows. hamy.xyz/blog/2026-06...

How I build vertical slices with agents: - Build slices of atomic commits e2e - Iterate to get desired behavior - Push bottom commits to stacked PRs - Regen stack based on feedback / improvements hamy.xyz/blog/2026-06...

How I Balance Speed, Quality, and Mental Overhead in Agentic Engineering - Drive strategy, systems, and vision - Let agents do the footwork quickly - Do 1-2 things at a time and improve the system each cycle hamy.xyz/blog/2026-06...

If agents are going to write the code anyway, why not pick the most reliable language? Python agents ship fast but fragile. Rust agents ship fast and correct - compiler catches ownership, lifetimes, and invalid states before runtime. Shift left is MORE important with agentic engineering.

QuietCut - a CLI that removes silence from videos: - Most of my videos are single take, talking head - This cuts editing time down to a single CLI call - Runs local, no subs hamy.xyz/notes/quietcut