Everyone hates on MCP because it adds a bunch of tokens to the context but like...that's not MCP's fault? Agents need to think about progressive disclosure everywhere and there's no reason that can't include both MCP server and tool activation.

It's funny how differently people use coding agents. I have never had to compact my history - I tend to break work up into tasks that can easily be completed without hitting context limits and spawn new threads frequently. But this seems to be an important feature to many!

Usually if I think of an "obvious" better way to do a software thing my priors are "well other people must have tried it and there's something I'm missing". I'm learning to reset that prior for working with LLMs. Tons of obvious stuff just lying around but yet to be tried.

2 best kinds of corp chats: - Chats with many people that become high-signal topic-based info sharing. Curated news better than any social feed. - Chats with <= 5 people that become extremely tight knit and candid over a long time horizon. Where career-long friends are made.

2 best kinds of corp chats: - Chats with many people that become high-signal topic-based info sharing. Curated news better than any social feed. - Chats with <= 5 people that become extremely tight knit and candid over a long time horizon. Where career-long friends are made.

A funny fact of tech is that almost every VC-backed startup has more resources at its disposal than almost every equivalent team inside a big tech co. There are exceptions for a few heavily concentrated bets, but in general the "David v. Goliath" dynamic isn't what you'd expect.

The Wax and Wayne books of the Mistborn Saga are basically genre catnip for me. It's a steampunk mashup of Batman, Sherlock Holmes, and Trigun and I'm enjoying them thoroughly while also feeling slightly embarrassed at being so thoroughly triangulated.

Bubbles are impacted both by difference between perceived and actual potential of a technology, and amount of uncertainty about winning applications thereof. Dotcom was a bubble where the perceived and actual potential was extremely high, but it took time to figure things out.

It should be legally required that patients can have access to digital copies of all medical records including imaging within 24h of requesting same Instead I'm overnighting a $6.50 check so they can burn a CD (I don't own a CD drive anymore so guess I'm buying one too).

"In the future apps are going away and AI will just generate interfaces on the fly." Have you seen the pitchforks that come out when popular apps make modest design changes?

Context Engineering is about managing the tension between NOT ENOUGH context (agent can't do task) and TOO MUCH context (agent gets off track). We're building complex RAG systems and endless bespoke tools, but we're sleeping on a powerful solution: the humble hyperlink. 🔗

No service is an island, Entire of itself. Each is a piece of production, A part of the web. Each service's outage diminishes me, For I am involved in prodkind. Therefore, send not to know For whom the pager beeps, It beeps for thee.

I think "politically orthogonal" movies are interesting...ones with an inherently political premise but that are not *about* politics. Civil War is about journalism. Eddington is about social disconnection. One Battle After Another is about parenthood. Hard to pull off!

Interactive coding agents work well for me, but trying to reliably build coding agent scripts is much less reliable. I think I don't even notice how often I'm Fonzie-ing the interactive agent - stopping it when it gets in a loop, abandoning an unsuccessful thread, etc.

If you own a product it is critical that you inoculate yourself against vanity metrics, stuff like "total users signed up" that can only ever go up. You want your product to succeed and the desperation to see positive signals is hard to overcome. But overcome it you must.

I find myself at a loss about AI video apps. Why would I want an entire app dedicated to lower effort content? I'm not angry it exists, I'm just confused why anyone would consume it. If the content is good enough to watch shouldn't it be able to compete on existing platforms?

Does *any* coding agent today have the ability to access MCP resources autonomously (via a built-in tool instead of by e.g. @-mentioning it)? There is such a huge opportunity for the first coding agent that successfully cracks intelligent context loading via resources.

Are you the author of an open source library and wish you had a better way to get coding agents to use it properly? DM me, would like to hear more. Working on something in this space.

Following AI right now is like playing competitive video games. When a new patch (model, tool, etc) drops, the balance shifts and new metas evolve (patterns of using various things together). And if you take a break for a few months you come back to a totally different game.

Feature request for browsers: keep track of what's under the mouse and when clicking if a layout shift changed target in last 500ms cancel the click event. Bug report for websites everywhere: Stop making interfaces where stuff loads in a way that causes these layout shifts.

I would be in favor of regulations that forced media platforms with kids' content have to have a mode where parents can strictly curate the entire library of available content. There's good educational stuff on Netflix/Disney+/HBO/YouTube for my daughter, but also total garbage.

One of my least favorite tropes is "kill the henchmen, spare the boss" where a hero battles through waves of enemies, fights a boss wins without killing and then a big deal is made of "sparing" the boss. What about the 37 people you murdered on the way in?

"Not Invented Here" (NIH) unfortunately applies not only between companies ("we won't use X because we will do a better job building it ourselves") but also within large companies ("we are the team who is supposed to own X so we won't use team Y's thing for it").

Hooks (e.g. Claude Code's hooks) are the critical primitive missing from the MCP spec. If an MCP server could hook into the agent lifecycle, a whole universe of possibilities opens up. At some point if I have time (so ~never 🙃) I intend to write up a formal proposal for it.

Pretty much every assistive AI system should allow for persisted customization of the system or initial user prompt. If you're trying to help the user do something, you need to let them define their own workflow guidance. This is the main reason coding agents are so popular imo.

Something that I think gets lost in the debate about whether 10x engineers exist or not is that there are multiple dimensions in play. I don't think there are engineers who would perform 10x better than the average engineer at *any* project/task. 2x? 3x? Maybe. But not 10x.

Want to bring 🤏🍌 into your app? Frontier image editing in Genkit is just a few lines of code away. Available now, no need to upgrade your Genkit version. Just use the 'gemini-2.5-flash-image-preview' model and set responseModalities in config.

Every website that does any kind of comparison between many options (flights, hotels, heck even shopping) should have an easy mechanism to save a list of potential options while searching and then compare just your saved options, picking one. But basically nobody does this.

Remote work isn't for everybody but I laugh at the idea of not being able to build friendships or community in a distributed team. Some of us grew up posting on Half-Life mod forums and posting pixel art on DeviantArt. Text-based community in-jokes and memes are in our blood.

Has anyone built a durable moat in the AI application layer yet? Like something so much better due to how it's put together that it doesn't matter if the underlying model is the best one. Coding agents answer seems no thus far -- people are switching on a dime to best model.

When I say "JS" I mean "also TypeScript" but when I say "JavaScript" I don't.

Is there a better pattern than the snippet below for "async initialization classes" where the constructor needs to kick off async work but I want to "push down" waiting for that async to the methods? I find myself using this pattern frequently but it feels clunky and ugly.

Vitest is an amazing test framework but it has an annoying flaw when building with AI - the default command starts a watch process which means you have to prompt around the default to let coding agents work with it.

It occurs to me that AI creates a more common need for two UX patterns: 1) Proposals: queue up a change to a resource that requires second party approval 2) Versioning: retain history for rollback and visibility Basically the systems we use to not trust humans, but for AI.

I'm fortunate in my career to get to work on things I really care about...because I literally cannot get my brain to work on problems I don't care about. You can tell me X is the most important thing for me to work on, but if I don't believe it I drop to ~20% capability.

Usually I'm afraid of studios learning the wrong lessons from box office (Barbie success? More toy movies!), but if Sinners and Weapons convinces studios to greenlight more auteur horror I'm all for it.

Pet peeve: systems that have monthly charts but not "last 28/30 days" charts. So annoying that I can only view the data in a useful form for previous months.

I think a huge number of devs will eventually have an itch to build their own coding agent that works just the way they want -- similar to web frameworks, bug trackers, etc. When a piece of tech becomes so central to your work, even small differences of opinion magnify.

I'm a dev but I'm way more annoyed by AI-generated writing than code. The purpose of code is its function - if the app works or the tests pass, I'm good. The purpose of writing is communicating - I don't want to spend time reading what you didn't spend time writing.

Imagining a sci-fi story where a far-future alien civilization tries to reconstruct what humanity was like based only on a thumb drive containing the weights to an LLM.

Tiny soapbox: "SWE-bench" is frustratingly misnamed as it consists *only* of Python code, and I genuinely think the name misleads people into thinking it's a general-purpose software engineering benchmark.

Using AI to enhance and scale human creativity makes sense to me, but fully autonomous media makes no sense to me. It's like saying the ultimate restaurant would be one where you can order whatever you want and a mediocre chef will microwave it from frozen ingredients.

Python has a problem with black magic -- libs hide critical details behind black box behaviors. Downloading models is the most obvious one. Models are often multi-gigabyte files and Python libs just auto-download them without so much as a Y/n prompt. Deeply unserious practice.

It's fascinating to me how OnTrac exists given they are materially worse at delivering packages than every other provider. Consistently late, consistently not following delivery instructions, and this has been true in multiple places I've lived.

Prediction: most services will largely migrate to remote HTTP MCP servers, but developer-oriented services will continue to ship stdio MCP servers for the foreseeable future. There's just too much value in having access to the local filesystem and shell.

Popular JS project: this has been tested on all LTS versions of Node.js, Bun, Deno, and 3 browser runtimes. It works with ESM and CJS and all popular bundlers. Popular ML project: this has worked exactly once on my machine running on a Python version released in 2022. Good luck!

One practice I learned in Go that I use everywhere now is to strongly prefer "table tests" where each test case is just a struct with various inputs and expected output states. It lowers the bar to adding additional tests as corner cases arise, and it's more readable.

Saddened to report that AI is no better at figuring out how to cleanly handle this merge conflict than I am. Long-lived branches are sometimes necessary but always painful.

HUGE update to Firebase Data Connect this month! 🔎 Full-text search support 🔢 Native enum support 💾 Get generated SDKs straight from console

I find I'm allergic to excessive positivity in business communication. If everything is 🌈🐶🦄✨ all the time then (a) when something actually amazing happens it doesn't stand out (b) you are probably ignoring your weaknesses I aim for a "grounded but slightly optimistic" tone.