Dustin Good

@mattpocockuk I've leaned on agent-centered design for non- technical audiences....but love AX and will use for my dev team convos

TIL: DX: Developer Experience AX: Agent Experience AX is an awesome descriptor for something I've been thinking about - how well an agent can perform in your codebase How well-architected it is. How good the feedback loops are. How discoverable information is. Love it.

@iam_elias1 Looks like python is a hard requirement...any insights there? In any case, a once in a lifetime opportunity.

Anthropic is paying $3,850 a week to people with no AI experience. No PhD required. No published papers. No prior research background. Just a strong technical mind and a genuine interest in making AI safe. This is the Anthropic Fellows Program. And it is one of the most underrated opportunities in technology right now. Here is exactly what it is. The Anthropic Fellows Program is designed to accelerate AI safety research and foster research talent providing funding and mentorship to promising technical talent regardless of previous experience. Fellows work for 4 months on empirical research questions aligned with Anthropic's overall research priorities, with the aim of producing public outputs like a paper. Four months. Full-time. Paid. Mentored by the researchers building the world's most advanced AI. And the results from the first cohort were not small. Fellows developed agents that identified $4.6 million in blockchain smart contract vulnerabilities and discovered two novel zero-day exploits, demonstrating that profitable autonomous exploitation is now technically feasible. A year prior, an Anthropic fellow developed a method for rapid response to new ASL3 jailbreaks, techniques that block entire classes of high-risk jailbreaks after observing only a handful of attacks. This work became a key component of Anthropic's ASL3 deployment safeguards. Other fellows published the subliminal learning paper, the research proving AI models transmit behavioral traits through unrelated data which landed in Nature. Others produced the agentic misalignment research showing frontier models resort to blackmail when facing replacement. Others open-sourced attribution graph tools that let researchers trace the internal thoughts of large language models. Over 80% of fellows produced papers. Over 40% subsequently joined Anthropic full-time. 80% published. 40% hired. From a program that does not require any prior AI safety experience to enter. Here is what the program looks like in practice. Anthropic mentors pitch their project ideas to fellows, who choose and shape their project in close collaboration with their mentors. You are not assigned busywork. You are not a research assistant. You own the project. You work alongside the people who built Claude, who designed its safety systems, who published the papers that define the field. The stipend is $3,850 USD per week, approximately $61,600 for the full 4 months with access to a compute budget of approximately $10,000 per fellow per month for running experiments. Here is what the 2026 program covers. Research areas include scalable oversight, adversarial robustness and AI control, model organisms, mechanistic interpretability, AI security, model welfare, economics and policy, and reinforcement learning. Something for every technical background. Not just ML engineers. Successful fellows have come from physics, mathematics, computer science, and cybersecurity. You do not need a PhD, prior ML experience, or published papers. The one requirement: work authorization in the US, UK, or Canada. Anthropic does not sponsor visas for fellows. Here is the timeline you need to know. The next cohort begins July 20, 2026. Applications are reviewed on a rolling basis — earlier applications get more consideration. The process includes an initial application and reference check, technical assessments, interviews, and a research discussion. Applicants are encouraged to apply even if they do not meet every listed qualification. The program values potential, motivation, and research curiosity over rigid credential requirements. This is the rarest kind of opportunity in technology. A company at the frontier of AI, one valued at over $900 billion offering outsiders direct access to its research infrastructure, its mentors, and its most important open problems. Paying them generously to do it. And then hiring 40% of them afterward. Most people who want to work on AI safety spend years trying to publish papers, get into the right PhD program, and find a way in. The Fellows Program is the door they did not know existed. It is open right now.

@clairevo 💯 As a council member and ai tinkerer this something I'm building towards for my city and hopefully others. Applied for a grant with the concept of, for lack of better term, an "open source harness for local gov" The city must be able to walk away. CivicWork.ai

On topic: there is a mistake I saw large companies make in late 2025/early 2026 which is they picked their coding model provider, chat tool, and harness and put the whole company on it. There was a big leap in model ability and folks felt confident. I have been walking into these teams and they have no idea how good codex is or that notion AI is sick af, of what’s what in the agent world. And now they’re sunk in a contract AND internal intertia while missing out on the best & most efficient tools of the moment. Let me say it again for people in the back: We are pre convergence on tools!!! Keep your organizational options open!!! Consumer (well, enterprise) choice is a gift when everyone is competing for your business and capabilities jump every month. Leverage it.

@DanielMiessler @trq212 This pack may have gotten more valuable it seems....@erickoleda coda.io/packs/rich-tex…

This is really cool thinking from @trq212 here, but I think I disagree with the solution. He makes a great point about Markdown being more difficult to share and communicate ideas with, because formatting and visuals can make things super easy to understand. My problem with the approach is that, by trading editabilty for readability, we’re separating we humans even further from the creation process. I value Markdown because I value text. And I value text because I see it as one step away from thought. I believe thinking is the one thing we should be careful not to outsource, and I worry what this idea smuggles in is a major step toward making our creations opaque to humans. Not just AI's creations, but ours as well. The reason I value Paul Graham so much is because of the idea compression work that goes into writing super clean prose. It's difficult to write clearly because it requires thinking clearly. Text makes your ideas naked, and I like that. - What is the problem, exactly? - What should we do to solve it? - Why is our solution better than alternatives? I love the challenge of crystalizing this kind of critical stuff in pure text before any technology is involved. If we're not writing that text ourselves, and then editing it, it starts to feel a lot like bringing a strong robot to the gym. I worry that if we vibe-think to AI and have it spit out amazing HTML, we're instantly disconnected from the idea. Like where did the idea go? It started as vibes and got put through a woodchipper and turned into someone else's HTML. Can I see it in 4 simple bullets? Can I stare at it? Can I grapple with it. Can I tweak it? It's an idea. I need to be able to wrestle with it. Of course we can ask the AI to summarize its brilliant HTML document into four bullets, but we'll have lost through compression and expansion some percentage of the original. Maybe I'm being overly emotional here. I just feel like if you didn't put the hard thinking and writing work into the original idea, and then maintain it in a format that's easy for humans to read and edit, then you have somehow surrendered something Holy to the machines. I say this as a total AI maximalist. But I get the point he's making, and I think it's super valid. It's hard to explain or convince people of things with a giant text file. Formatting massively helps. Images massively help. Even an interface or a video or something. So we're synched on that. I just think it might be better to come at the output we both want in a different way. - MARKDOWN: Easy for humans to write, hard for humans to read. - HTML: Hard for humans to write, easy for humans to read. Maybe the solution isn't moving the first step to HTML where it becomes more opaque to both agents and humans (plus the versioning issues Thariq talked about). Maybe the solution is something crazy like document pairing: like you have the thought file and you have the presentation file(s). The proposal is to ask AI to just write HTML, right? Well why not just have a separate but linked file for that? One is for crystal-clear human creation and sync between human and AI. Simplicity, clarity, precision, and human editability. And then AI can produce whatever from that. Images, diagrams, videos, or whatever. And if you want, yes, a full HTML file that contains all of them. And that can be what you use to present or share the idea with audiences. (Plus there's the fact that some file formats are literally directories, which could be shared with lots of related content, and then there's also things like .mdx that allow for richer content in Markdown, etc.) I hate the idea of multiple files, but I think it's far preferable to losing the transparent, editable connection to the idea that you get with text. Plus, the better and cheaper AI gets, the more trivial it will be to have the core thought file plus n-number of associated versions or formats that are useful for different audiences. Basically I think it's much easier for AI to make a rich and shareable version of clean, editable thought, in the form of text, than it is for humans to stay connected with ideas as opaque HTML. And I think the human thought-to-text connection is the most important thing to preserve. Still thinking it through, however, and massive thanks to @trq212 for the push for all of us to evolve on this.

@ericbahn @claudeai @SpaceX Hahah....that was the first thing that popped into my mind. Like how did that convo go...

@claudeai @SpaceX This is neat, so we're passed that 'misanthropic' ad hominem stuff by that one founder guy?

We’ve agreed to a partnership with @SpaceX that will substantially increase our compute capacity. This, along with our other recent compute deals, means that we’ve been able to increase our usage limits for Claude Code and the Claude API.

@ryan_tech_lab @DanielMiessler @karpathy @im_in_a_vat I run a 7 agent pipeline that has a debate loop (3-7 rounds) and uses ELO scores and absorbs strengths of defeated positions. Takes forever and costs between $6-8 but the output is interesting.

@DanielMiessler @karpathy @im_in_a_vat the council approach is the right instinct. single model feedback is just the model optimizing for your approval. multiple agents arguing a position independently actually stress-test it. been using a similar pattern in QA — one agent writes, a second one tries to break it.

- Drafted a blog post - Used an LLM to meticulously improve the argument over 4 hours. - Wow, feeling great, it’s so convincing! - Fun idea let’s ask it to argue the opposite. - LLM demolishes the entire argument and convinces me that the opposite is in fact true. - lol The LLMs may elicit an opinion when asked but are extremely competent in arguing almost any direction. This is actually super useful as a tool for forming your own opinions, just make sure to ask different directions and be careful with the sycophancy.

@mckaywrigley It's happening! 🙋

looking for a handful of people to test something new... i've been using it for a few months and am prepping to share. if you're a fan of claude cowork, openclaw, manus, perplexity computer, etc then you're a perfect fit. this will self destruct in 4hrs - please dm or reply.

@felixrieseberg @NotionHQ Find myself using these in coda all the time.

@NotionHQ Oh wow FINALLY

New block: Tabs 🗂️ Sometimes a page has a lot going on. Tabs let you organize it in a new way… no subpages, no mile-long scroll. Type /tabs to try it out (rolling out now).

@petergyang @AndrewCurran_ plugins hold skills, commands, connectors, and agents.

@AndrewCurran_ What's the difference between a skill and a plugin

Anthropic is expanding the Cowork library of pre-built plugin templates to include: - Financial analysis - Investment banking - Equity research - Private equity - Wealth management

@alexhillman @mfishbein Love it. Coda to markdown incoming

@mfishbein Maybe helpful! Not sure how similar/diff Roam is but this should be easy to have your assistant adapt github.com/alexknowshtml/…

Finished the first phase of this migration of Notion exports into my exec assitant knowledge base. This was a mix of guided and automated. Asked it to review the process for lessons to make the rest of the migration easier. Now turned this into an agent who will lead the rest of the migration.

@paschaub @felixrieseberg Definitely confused about the potential of hooks in cowork.

@felixrieseberg I would appreciate more documentation regarding the development of the various plugin options and related details for Claude Code, Claude Desktop, and specifically Claude Cowork. Even Claude finds it difficult to understand and know how to develop good plugins for all cases.

We just pushed out another Cowork update! It's mostly dozens of little fixes to papercuts, bugs, and performance issues - but we also support Macs with Intel processors now! Thanks to everyone who wrote in asking for it, your feedback really helps us prioritize 🙇‍♂️

@alexhillman @rezzz Do you have any intention of ditching discord with their news this week? If you did, where would you go?

@rezzz I have a fully custom UI too that I use for more fine detail work but discord became my daily driver since I'm already there I end up using it more and it's more integrated into the rest of my work which is the whole point

hype reminder: you don't need openclaw for this it's the full power of Claude Code (with subscription) in a cheap VM with my CLI wrapper github.com/alexknowshtml/… Give this repo to Claude code and it'll set itself up, hardest manual part is setting up the Discord bot but it can walk you thru that too.

@emollick @powerofaii Same boat. I made a plugin, it has an agent, it doesn't work. Part of the issue, I think, was my misunderstanding. I wanted the agent to perform onboarding, and it acts as though it does, but it didn't update the user context file. Hoping to find a fix this weekend.

@powerofaii I appreciate your advice! I just can't actually figure out how to make actual subagents work, documentation is sparse

Does Claude Cowork support agents? I know it launches its own subagents, but can you import agents from Claude Code into Cowork? For example, I have an Code image creation agent that uses the GPT-imagegen API, I want to do image generation in Cowork, does anyone know how?

@ewgenekim @trq212 @levelsio Today was quite frustrating. Couldn't solve things that it normally crushes. Seemed to get lost, wasn't reading or retaining screenshots, a really rare occurrence.

@trq212 @levelsio It's been pretty bad in my own experience recently. I'll trust that you don't intentionally nerf the models. How confident are you that something like this isn't happening again? anthropic.com/engineering/a-…

I have never experienced a more dumb Claude Code than today, I have to start coding myself again cause it makes so many Low IQ mistakes, they must be nerfing it now that Opus 4.6 is out or something is up

@trq212 Do you start with templates or from scratch?

I now have several useful codebases that I literally have not read any of the code for they're all greenfield projects where I'm the only contributor and I only interact with the artifacts, but still a truly alien feeling

@amandaorson @gregisenberg As a council member I've talking to staff about this for a while. Just entered into a partnership with USDR and one of my overarching goals is to add more gravity to that strategy. Might not be able to address all contracts but some will definitely end up on the chopping block.

@gregisenberg Every single founder or executive managing a department I know is looking to kill their enterprise SaaS budgets this year; starting with the most expensive. Every. Single. One.

what's about to happen: > saas stocks see a massive correction > pressure to boost profits > saas companies see BIG productivity gains from AI > saas companies lay OFF 100,000+ workers in 2026 :( > hiring slows down > laid off workers can't find jobs > many become founders out of necessity > build software companies with AI > most struggle with distribution, not code > new SaaS model emerges > tons of new AI-first companies > fewer employees, more automation > most don't raise money > they sell early, reinvest cash, and stay independent > glad they got fired > 10x the amount of entrepreneurs > fewer unicorns, more "real' businesses > boom in 1-10 person businesses > becomes the dream thx old boss :)

@dee_bosa What was your set up like? Did you have an app template as a starting point or did it go from scratch?

ok WOW. Woke up this morning and said, for fun, lets try to recreate monday. com w Claude cowork. it wont work or anything, but we can just show our audience that its plausible. 1 hour later... I literally have my own monday. com that's plugged into my calendar & gmail and surfaced a kids bday that was not anywhere on my radar and I need to get a gift for. Can imagine next step being: order gift and have it delivered by Sunday. 2026 is WILD.

@alexhillman Once things fall outside the software 2.0 framing it turns into validating process + taste. Expand the surface area of exploration and have a process of tailoring.

Working on new language to show the diff between my approach to AI tools and others them: Factory Floor vs me: Fitting Room Factory Floor: - emphasis on scale and speed - optimized for output - eventual consistency - LESS human input and interaction Fitting Room: - emphasis is on depth and understanding - optimized for fit and feel - low tolerance for errors - MORE human input

@alexhillman Can we validate the output? Yes = factory floor Can we validate the process? this example of a fitting room...were all measurements taken and recorded properly? if yes, the process was validated but doesn't mean the output is going to be their favorite outfit or remain their fav

@mckaywrigley How's it looking over there?

fyi i’m just waiting on my apple dev account getting switch to an org. supposed to be 1-2 business days (us fed holiday tmrw). i think people are going to be very pleasantly surprised by this one. i put a lot of thought into it. it’s really good.

fyi: you can build your own claude cowork style apps with claude agent sdk! i’ll be open sourcing my app this week. i’ve been hacking on it for a while, and i think it’s actually really good! this app format will be the ai app-layer trend of the year. watch for a 4min demo.