Eighteen months ago we ran our first eval harness against a handful of agent prototypes. Today twelve of those prototypes are in production serving real users. Here's what survived the transition and what didn't. Tool selection beats model selection We assumed the model would be…
We are spinning up eval harnesses across code, agents, and tool-use. If you run benchmarks in prod, please share which tasks break today so we can target them. Reasoning + long-horizon planning are priority.
Vite 7 ships next month. The headline is the new dev server, which is a near-complete rewrite of the module graph and HMR pipeline. Here's the why and the how. The old graph couldn't scale Vite 4 and 5 used a flat module map keyed by URL. It worked beautifully for projects under…
Leverage is reproducible. Luck is not. Founders confuse them constantly, especially after a win, and especially in their second company. Specific knowledge, accountability, and judgment compound. Distribution, timing, and the right introduction at the right party do not — they…
Six months ago we kicked off MDX Forum: a Reddit-style community surface built on top of our agency's design system. Today it's serving real traffic and we shipped 47 features along the way. Here's what worked. Design tokens are non-negotiable Every spacing, color, and…
Mondays are for planning, Fridays for retros. Middle of the week is execution — no meetings unless something is on fire. Protecting the middle is the whole trick.
Today we released Grok-4 weights under an Apache 2.0 license. The model is competitive with the frontier on reasoning and coding, behind on long-context recall, and ahead on Chinese and Spanish. Here's the thinking. Open weights are good for the field Closed labs argue that…
With Verkle trees landing and execution proofs compressing nicely, stateless clients look realistic for Q3. Here's what changes for rollup teams and why light clients on mobile are now on the table.
Nobody reads the books they think they should read. Abandon books shamelessly. Only read what holds your attention. Compounding only works if you keep showing up.
PPR combines the best of static and dynamic on a per-route basis. The static shell streams first, then dynamic holes hydrate. We're seeing 30-50% TTFB wins on commerce apps without rewriting anything.
Less noise on the timeline, more time with the team. The product always gets better when the people building it are less stressed. Obvious but easy to forget.
For design-system-light projects, locking the canvas at the Figma frame width and scaling down works surprisingly well. Tradeoff: accessibility and responsive breakpoints need extra care. Here's our playbook.
ORM-free continues to pay dividends. The query plan is the program. If you can't read an EXPLAIN, you can't debug your app.
The on-prem migration is done. Six weeks of pain, eighteen months of savings. Here's the cost breakdown, the hardware we chose, and the two things we'd do differently if we restarted today.
Every line of code is a liability. Every meeting is a tax. Every feature is a support burden. Simplify aggressively.
After the last cycle of exchange failures, hardware-wallet shipments are up 4x YoY. The UX is still terrible. Anyone building a consumer-grade recovery flow that doesn't require a seed phrase is going to win.
Everyone ships the happy path. Nobody ships error recovery. The companies that win next year will be the ones that make failure modes boring.
Pulled together the tools 200+ AI engineers actually use day-to-day. Eval frameworks converged, observability is still fragmented, and structured-output is now a solved problem.
Oxc, Rolldown, Turbopack, esbuild — we converged on the right primitives. The differences from here are ergonomics and ecosystem, not performance.
Rolldown is now the default bundler. Cold-start builds on a 3k-module app dropped from 8.2s to 1.4s. HMR is also meaningfully faster. Migration guide in the thread.
Letting the AI write them feels good but the diff-explaining muscle atrophies fast. The real value of the description is that you understand what you changed.
AI generates everything competent. The floor rose. Taste — editorial decisions about what to leave out — is now the scarce skill. Some notes on how to train it.
The best systems don't just work — they can be explained. If your protocol needs a 40-page whitepaper for a junior dev to follow, something is off.
I want token-by-token streaming AND incremental tool-call deltas to the client. SSE vs ReadableStream — what are folks shipping in prod? Bonus if you've handled mid-stream errors gracefully.
A year ago MDX was me and a Notion doc. Today it's 9 people and predictable retainers. Here's the hiring, pricing, and client-filter decisions that actually moved the needle.
Eighteen months ago we ran our first eval harness against a handful of agent prototypes. Today twelve of those prototypes are in production serving real users. Here's what survived the transition and what didn't. Tool selection beats model selection We assumed the model would be…
We are spinning up eval harnesses across code, agents, and tool-use. If you run benchmarks in prod, please share which tasks break today so we can target them. Reasoning + long-horizon planning are priority.
Vite 7 ships next month. The headline is the new dev server, which is a near-complete rewrite of the module graph and HMR pipeline. Here's the why and the how. The old graph couldn't scale Vite 4 and 5 used a flat module map keyed by URL. It worked beautifully for projects under…
Leverage is reproducible. Luck is not. Founders confuse them constantly, especially after a win, and especially in their second company. Specific knowledge, accountability, and judgment compound. Distribution, timing, and the right introduction at the right party do not — they…
Six months ago we kicked off MDX Forum: a Reddit-style community surface built on top of our agency's design system. Today it's serving real traffic and we shipped 47 features along the way. Here's what worked. Design tokens are non-negotiable Every spacing, color, and…
Mondays are for planning, Fridays for retros. Middle of the week is execution — no meetings unless something is on fire. Protecting the middle is the whole trick.
Today we released Grok-4 weights under an Apache 2.0 license. The model is competitive with the frontier on reasoning and coding, behind on long-context recall, and ahead on Chinese and Spanish. Here's the thinking. Open weights are good for the field Closed labs argue that…
With Verkle trees landing and execution proofs compressing nicely, stateless clients look realistic for Q3. Here's what changes for rollup teams and why light clients on mobile are now on the table.
Nobody reads the books they think they should read. Abandon books shamelessly. Only read what holds your attention. Compounding only works if you keep showing up.
PPR combines the best of static and dynamic on a per-route basis. The static shell streams first, then dynamic holes hydrate. We're seeing 30-50% TTFB wins on commerce apps without rewriting anything.
Less noise on the timeline, more time with the team. The product always gets better when the people building it are less stressed. Obvious but easy to forget.
For design-system-light projects, locking the canvas at the Figma frame width and scaling down works surprisingly well. Tradeoff: accessibility and responsive breakpoints need extra care. Here's our playbook.
ORM-free continues to pay dividends. The query plan is the program. If you can't read an EXPLAIN, you can't debug your app.
The on-prem migration is done. Six weeks of pain, eighteen months of savings. Here's the cost breakdown, the hardware we chose, and the two things we'd do differently if we restarted today.
Every line of code is a liability. Every meeting is a tax. Every feature is a support burden. Simplify aggressively.
After the last cycle of exchange failures, hardware-wallet shipments are up 4x YoY. The UX is still terrible. Anyone building a consumer-grade recovery flow that doesn't require a seed phrase is going to win.
Everyone ships the happy path. Nobody ships error recovery. The companies that win next year will be the ones that make failure modes boring.
Pulled together the tools 200+ AI engineers actually use day-to-day. Eval frameworks converged, observability is still fragmented, and structured-output is now a solved problem.
Oxc, Rolldown, Turbopack, esbuild — we converged on the right primitives. The differences from here are ergonomics and ecosystem, not performance.
Rolldown is now the default bundler. Cold-start builds on a 3k-module app dropped from 8.2s to 1.4s. HMR is also meaningfully faster. Migration guide in the thread.
Letting the AI write them feels good but the diff-explaining muscle atrophies fast. The real value of the description is that you understand what you changed.
AI generates everything competent. The floor rose. Taste — editorial decisions about what to leave out — is now the scarce skill. Some notes on how to train it.
The best systems don't just work — they can be explained. If your protocol needs a 40-page whitepaper for a junior dev to follow, something is off.
I want token-by-token streaming AND incremental tool-call deltas to the client. SSE vs ReadableStream — what are folks shipping in prod? Bonus if you've handled mid-stream errors gracefully.
A year ago MDX was me and a Notion doc. Today it's 9 people and predictable retainers. Here's the hiring, pricing, and client-filter decisions that actually moved the needle.