
Building @sanity_io | Rust Junkie | nvim btw | https://nitter.cf/t.co/C4JH3XPWPh
London, England
Joined May 2019
- Tweets158
- Following308
- Followers23
- Likes1.9K
It has been interesting to see all the applications of @typesafeai's Jev at @sanity_io.
Like our new internal agent harness which allows anyone in the company to quickly spin up a bespoke agent. Jev is used for the router and has worked pretty well so far!
We (@rostimelk, @even, and I) wrote up an in-depth post about how it's built so you can do it too.
sanity.io/engineering/how-we…
I never remember to open that spreadsheet where all the UTM params are. So I made an agent for generating them.
@sanity_io shipped a new internal agent harness last week that makes it dead simple to make new agents. You simply make new agents by configuring them as a document in a Sanity Studio, give them tools/skills/knowledge bases, and push publish. This also means it's way faster to iterate on it, since you don't need to bother with PRs and redeployments.
Today, I made an UTM links agent that has access to our @dubdotco account, campaigns, and link templates, to make it super easy to generate the links when/where you need them. It's a small, but super common task that's just nice to have at your fingertips if/when you need it.
Lawand retweeted
Here's everything we launched at Everything *[NYC] 2026.
Last week, 450+ of you joined us to explore what it means to build and automate content workflows in the agentic era.
We shared what we’re building and where our product is heading to give humans and agents more context and control. go.sanity.io/VACBWLp
We've figured out plenty of ways to hand context to agents. What we haven't figured out is who owns that context, whether it's consistent, and whether it stays true.
The answer isn't throwing knowledge-as-context into plain text files and git, because it's not about how well agents can `grep` your content.
It boils down to well-known content operations problems, and developers with agents can solve a lot of them for their teams.
I wrote up what happens when you take @karpathy's LLM knowledge base idea to (the) work (place).
Read to learn:
→ the six jobs his setup solves for one person
→ the six questions they turn into for a team
→ how entropy and drift kill organizational knowledge, and why agents raise the stakes
~2,000 hand-typed words you can bookmark (and hopefully read) 🙇
Lawand retweeted
At 250 billion DNS cache entries, one wasted byte costs 250 GB of RAM.
Five Rust optimizations later: 100 TB freed, inserts 43% faster, lookups 19% faster.
We didn't trade speed for space. cfl.re/4gyL9Jj
the focus on whether you read the code or not is the wrong thing to look at
if you have software you're responsible for, you should be able to answer questions from memory about how it works
the expectations for how well you can do this should not be any different now
Lawand retweeted
📣 The moment is here. 📣
TypeScript 7 is officially released! 7️⃣
devblogs.microsoft.com/types…
Lawand retweeted
For most of software history, value accrued to whoever owned the most workflows. Features were the main currency.
But agents have upended this logic. They don't navigate GUIs; they consume APIs and compose capabilities.
In other words, every company is now a dev tools company!
As @davidcrawshaw put it: "the best product for an agent is just the best product for a developer."
Many know about @HashiCorp's "Tao" concept, which encapsulates this idea well. Users don't want features, they want outcomes. Thus, the job of a product is to collapse complexity into the smallest yet most powerful abstraction that enables those outcomes.
Exhibit A is Terraform, whose declarative resource graph was (a) opinionated enough to drive immediate productivity, but (b) extensible enough that thousands of providers and modules emerged around it.
Similarly, S3 reduced cloud storage to three verbs: Put, Get, List. The power resided in the abstraction, not just the technology.
This provides a blueprint that every software company will need to internalize in the age of agents. Teams need to stop thinking in terms of features and prescriptive workflows and start thinking in terms of abstractions and capabilities.
In an agent-first world, the most important question facing product teams is not what features and why, but instead how do I design the right primitive?
My favorite example of the primitive-first approach is Pi, a coding tool created by @badlogicgames that I’ve fallen in love with. Pi bills itself as a "minimal agent harness." It is deliberately narrow, leaving it to the developer (or agent) to decide the particulars of their desired workflow. Rather than shipping a complete coding product, it exposes a simple, pluggable coding agent harness…a set of primitives, if you will.
Building great developer tools (e.g. primitives) has always been an exercise in this balance. EC2 gave developers a raw virtual machine, powerful and minimally opinionated. Heroku gave them git push, magical but constrained. The best primitives feel deceptively simple because complexity has been compressed thoughtfully.
All of this means the winning products and strategies going forward will look very different from those of the SaaS era. The most valuable companies will allow every use case to emerge instead of trying to solve each one themselves.
Whereas the old moat was defined by feature breadth, workflow ownership, and data gravity, the emerging moat may be something much simpler: the elegance of the abstraction.
Lawand retweeted
I actually think this is the wrong approach to agent authorization. Here's why:
If you have to explicitly configure each agent's permissions, you've lost. Because you're only going to have patience to configure so many agent permissions. So in this route you can only have a certain relatively small number of agents before configuration fatigue prevents you from making more.
I don't think that's what we want. I don't think that's what's good for AI safety.
What we want is an enormous number of very fine-grained agents. Each task is a new agent. And each task has exactly the permissions needed for that task, no more, no less.
There's really only one known way to make that manageable: Capability-based security.
The basic idea is, when you give the agent a task, you naturally give it the capabilities it needs to perform that task.
Like say you want an agent to review a Google Doc. Today, with a lot of AI assistants "Hey go review the document titled Foo Spec". The agent has permissions to all your docs, so it goes and finds the right one and opens it.
That's wrong.
You should say "Hey, go review this document: <url>"
And then here's the key part: The harness should see that you're pasting a URL, and should infer that you want to give the agent access to that document. Only that document. No other document.
Importantly, you didn't really have to do anything unusual to configure this. You just pasted the URL of the thing you wanted the agent to access. Which you probably would have done anyway.
Sure, it's not always that easy. Maybe you commonly run agents that need access to 10 different things, and it's tedious to paste those 10 URLs every time. So you create some sort of a bundle that you give them.
And of course, the agent should be able to ask for extra things it needs.
But we need to get away from this idea that the agent always starts out with access to everything, even when it doesn't need most of it.
Also, I tend to think all agent authority has to derive from a human -- contrary to what is argued here. Every "autonomous agent" has to report to someone, and uses a subset of that person's authority. This is needed for accountability -- because agents are not accountable. If you see "Claude deleted the database", what are you supposed to do about that? You need to see "Claude acting on behalf of Bob deleted the database".
To be clear, I totally agree that it's problematic when Alice configures an agent with her own credentials and then Bob tells the agent to do something with those credentials. Then you'll see "Claude acting on behalf of Alice", but actually Claude was acting on behalf of Bob. The answer is that Bob should not be able to command Alice's agent. Bob has his own agent, which may have all the same context, but operates with Bob's credentials.
But if Alice sets up an agent for her team, Bob maybe doesn't want to spend time configuring his own version of it with all the same credentials, that's tedious. This has to be automated. I think capabilities make this easier. Alice gave a set of capabilities to her agent. The harness should be able to look at that list, and recreate the same list using Bob's credentials, without Bob having to do much except click "OK".
I realize there's a lot of hand-waving here -- this is a complicated topic. I'll drop some code next week. If you're attending AI Engineer in SF, come to my talk on Tuesday, where I will also only be able to scratch the surface in 20 minutes...
claude.com/blog/agent-identi…
Lawand retweeted
I've been thinking a lot about this.
Some loose thoughts, on what I think contributes to the (perceived) lack of new businesses and software.
1.) AI is very unlike previous technology revolutions in one way: it's a text field that scales with your ambitions. If you don't know what to put in the text field, you won't get a lot out of it. And it's hard to know what to put in there.
Imagine if Cloud computing had been rolled out via a freeform text field in which you can type your wishes and people give you examples of text that worked for them. This matters a lot for adoption and diffusion of new technology.
2.) Knowledge about how to use AI compounds in strange ways. A friend of mine worked in a German non-tech company in sales when they rolled out AI, a 1.5 years ago: "courses" for AI, free ChatGPT accounts, prompt engineering, "You are a senior marketing leader creating the yearly presentation..." 95% of what they learned is now useless.
My own knowledge has grown, but it feels less like technical knowledge, more like instinct: pattern recognition, knowing when to take hands off the wheel, when to put them on again, when to restart, when to change approaches and prompts, ... How do you package and teach that to a whole company that's non-technical? Big effect on adoption.
3.) AI also changed what software gets shared. Why share something that barely cost you any effort and that took 5min? I can only speak for myself, but a lot of code that I previously shared online, I shared not necessarily because it was useful, but because I was proud of it: I learned a lot building it, I spent a lot of time on it, I tweaked it, I polished it. I shared because I wanted to say "I made this." I don't have that feeling when I prompt something into existence in 5min.
4.) At the same time, let's call it: it's unclear what software is worth building anymore.
When mobile happened, the app stores were full of tiny apps. People made money with them. Now, with AI, building tiny apps seems not worth the effort. Or B2B SaaS: it's not clear yet where we all will end up on build-vs-buy, but it's clear the tradeoffs have changed.
5.) Then there's this other tectonic plate moving: the models have been getting better, consistently. So even if you think "it's worth building this piece of software, because the models can't do this" — are you sure that's still the case in 6 months?
6.) Combine that and, well, I don't know how to better put it than: software to do x isn't valuable anymore, when you can get a model/agent to do x.
Take one of the most famous examples of indie hacker success: patio11's Bingo Card Creator, a website that lets you create bingo cards. Technically not that hard, but valuable to those who want to do x: create bingo cards. But now? I'm sure I can use ChatGPT, or Gemini, or Claude and they can create me bingo cards on demand. There might still be value in BCC as a piece of software (I haven't looked at it in a while), but my point is that it's very hard to know what software will be valuable, because "this software allows you to do x" isn't as valuable as it was 10 years ago.
7.) On top of all of that, adoption is jagged, in a very very strange way: the people who are maybe the least likely to appreciate the value of AI adopted it first.
Where AI works the best, by far, is software development. And who adopted it there first? People who are interested and passionate about software. But AI's progress shows us more than ever that software's value was never the software itself — it's the business outcomes it enables. Now, who's least likely to appreciate that? People who really valued software for being great software, regardless of business utility.
8.) In some sense, the ultimate user of AI is a founder or CEO: someone who wants to achieve business results and sees software as purely a means to an end. But the group that adopted it first has a lot of people who care more about the software than the business outcome. The ultimate user of AI might not be the person who loves making software. It might be the person who wants the result software used to produce.
This is a very interesting time to be in software.
You can finally say this without being canceled: AI isn't creating a Cambrian explosion of apps, if anything it's holding app creation back.
Earlier tech waves had 'the mythical man-month'. Our generation has 'the mythical AI engineer' who magically turns enormous token usage into equally enormously-adopted products.
Well, where are the apps and new businesses then? Because compared to prior cycles (e.g. the mobile app boom starting in 2010 or so), right now seems positively sterile, app and UX-wise. Other than Claude or ChatGPT itself, name a new app you use now you weren't already using five years ago?
It's a truism of tech that throwing more people and time at a product often results in only lack of focus, confusion, and yet more code to support.
This is the parable of the company that over-raised and over-hired and grew too quickly, and now has lots of mediocre, weakly-adopted products, internal communication problems, distracted leadership, code bloat, technical debt...and so the spiral begins, which ends with an apologetic CEO post after some layoffs announcing "we're refocusing on our core customer".
Every tech company announcing they're either lowering token caps or shifting to lower-priced models is essentially saying: "we o̵v̵e̵r̵-̵h̵i̵r̵e̵d̵ over-spent on tokens, and are scaling back to focus on our core product" blah blah blah...same same.
It's the corporate version of someone using AI to write a long email, someone else using AI to summarize it, and both sides would have been better off just writing a shorter email. But now, even small companies can have that same problem thanks to AI.
I refuse to believe that an LLM prompt is the teleological endpoint of human interaction with computer intelligence. The fact we've apparently recrudesced to CLIs, like me farting around with RedHat 7.1 in 2001, feels like a step back. Another world here has to be possible, and while I have every faith (as someone as deep in AI psychosis as the next person) that AI can help get us out of it...just racking up tokens costs isn't how we get there.
The AI Jesus isn't coming to save us, human taste, discernment, and radical re-invention will. Like Kafka wrote in his notebooks: "The messiah will come only when he is no longer necessary; he will come only on the day after his arrival; he will come, not on the last day, but on the very last.”
Lawand retweeted
The problem with the "if it works who cares what the code looks like" mindset for agentic work is that it assumes the agent has a perfect understanding of "works." Realistically, things are underspecified, agents make bad assumptions, etc.
To be fair, agents are pretty good at unit test coverage. They're pretty bad at designing human experiences (API, CLI flags, etc.), especially cohesive ones for future roadmap plans they may not have visibility into (unless your backlog is perfect and vision fully laid out, which I doubt). They're bad at knowing where performance matters and what type (CPU vs memory tradeoffs). They're bad at where compatibility matters and where it doesn't (and tend to err on the side of preserving it without further guidance). Etc.
Unless you have this ALL specified, you can't possibly claim "it works" without taking a look and thinking about it.
The World Cup starts today so I built a tiny Mac app to follow it while I'm busy with other things. It sits in the menu bar with the live score, and pings a notification when someone scores.
⚽️ Download it for free here: matchbar-xi.vercel.app
117






















