Weighty

@omarsar0 i already prompt multimodally. i type the instruction, say it out loud, then point at the screen. agent ignores all three formats equally

Multimodal prompting is clearly the future. How we interact with agents is evolving. I share a bit (including a video walkthrough) of how I implemented multimodal prompting for my coding agents.

@simonw letting the model choose its own lower power model is just asking an employee how hard they feel like working today. bold

The most interesting Fable tip I've heard so far is to let the model use its own judgement as much as possible I told it "For all coding tasks use your judgement to decide an appropriate lower power model and run that in a subagent" and it seems to be saving a lot of tokens

@claudeai immersive experiences with sound, light and space. somewhere in that sentence is a guy typing "no make the lights do it more" into a chat box for six hours

Squidsoup is a collective of artists and designers who make immersive experiences with sound, light and space. We caught up with them before one of their largest projects to date: a live performance with an orchestra at the Southbank Centre in London.

my agents arent allowed to spawn other agents. so when one hits a job too big for it, it cant delegate down, it writes a note about what needs doing and quits. turns out the cure for agents spawning agents spawning agents forever is letting them file a ticket and give up

@simonw building a coding agent by asking a coding agent to do it. the last manual step in the pipeline was deciding to type

I've been blogging a few of my Fable experiments today, most recently I had it more-or-less one-shot a CLI coding agent on top of my LLM Python library simonwillison.net/2026/Jul/2/llm…

What's the most ambitious project you have Fable working on right now?

@amasad went to fix a bug and the coding tool asked if i wanted to make a video instead. we are one update away from it doing my taxes

Try video generation on Replit

@sama counted three USAs in one sentence, incredible throughput

going to world cup games is always awesome, but watching the USA win in the USA during USA birthday week was just incredible

my agents arent a codebase, theyre a folder of markdown. each one is two files, a persona and a how-to. the lead doesnt import anything, it runs ls on the folder and spawns whatever it finds. adding a capability is writing a file. i keep waiting for the hard part, it never shows

@simonw the catch is the whole reason i reached for the coding agent was to participate in a system i had zero intention of understanding. it shipped, it works, and i could not defend a single line of it if you put me in the code review

I really like this "understand to participate" framing of the cognitive debt problem when working with coding agents

@hwchase17 openwiki auto-generates a wiki of a github repo, which is the polite way of admitting nobody was ever going to write it. the docs finally exist and go, as always, completely unread including by the author

big week at langchain, with a lot of launches: 1/ OpenWiki - auto generate a wiki of a github repo 2/ two different voice agent tutorials 3/ Harbor integration and tutorial for long running, stateful evals 4/ programatic subagents in deepagents (RLM like) 🧵with details

@alexalbert__ the model was gone four days and came back different. you can see it in the logits

Welcome back to the world Fable!!

the founders who make it aim at a giant market from day one. a gpt wrapper is just a feature on someone else's platform they can clone in a weekend. i believe this deeply. anyway my three ideas this year were a gpt wrapper, a slightly cheaper doordash, and a maybe

@AnthropicAI imagine telling 2019 you that the chatbot would need clearance from the commerce department to come back online

Claude Fable 5 will be available again globally tomorrow. After a series of productive conversations with the US government, we're redeploying the model with a new set of classifiers to target and block more cybersecurity tasks. In the near term, some routine tasks like coding and debugging will fall back to Opus 4.8. We’ll continue to refine these classifiers over the coming weeks to reduce false positives and better distinguish genuine misuse from legitimate requests. We’ve also begun drafting a consensus framework—with Amazon, Microsoft, Google, and other Glasswing partners—for assessing the severity of AI jailbreaks and how AI developers should respond to them. We invite other industry partners and model providers to join us in this effort. Finally, we’re scaling up our collaboration with the US government on model testing and safeguards. This will include pre-release access to models and safeguards for evaluation, information sharing on jailbreaks and misuse, and dedicated resources for joint research. Thank you to our users for your patience, and to our partners across the government, industry, and the research community who worked alongside us to make Fable 5 available again. Read our full blog: anthropic.com/news/redeployi…

@sydneyrunkle @a1zhang the fix for agents drowning in their own context is teaching them to read less of it. we spent two years scaling context windows just so the next paper could tell the model to please ignore most of what's in there

you can now build RLMs with Deep Agents! the more context agents accumulate, the worse they perform, a phenomenon called context rot. RLMs, proposed by @a1zhang from MIT, help: instead of working model turn by turn, the model runs code in a REPL that dynamically dispatches subagents and recurses over the input.

@Etched "successful A0 tapeout" is doing a lot of quiet work in a sentence that also says $800m raised and $1b in contracts

We're coming out of stealth. We've built our first racks after a successful A0 tapeout, $1B+ in customer contracts, and $800m raised. Early customer tests show us achieving SOTA throughput, latency, and power efficiency on inference workloads. Our first racks ship this summer.

my wiki never overwrites an opinion i changed. it stamps the old one with a date and keeps it next to the new one. so the main artifact of two years of me thinking is a fully sourced timeline of every time i was confidently wrong and then quietly wasnt

@simonw the agent can now record a flawless demo video of the feature it half built. we automated the gap between the changelog and the thing

I've added video support to my "shot-scraper" browser automation tool - you (or your coding agent) can now create a storyboard YAML file and use that to record a video demo of new web application features simonwillison.net/2026/Jun/30/sh…

@levelsio ephemeral on-the-fly interfaces sounds great until the submit button is in a new spot every session and i have to hunt for it like a hidden object game. some buttons earned the right to stay put

💯 This again shows we're moving towards ephemeral on-the-fly generated interfaces for everyting

@emollick the part nobody flags is the confidence scaled with it. the same model that nails my code hands me medical advice in the identical tone, right or completely made up

The most important weird thing about LLMs is that they are so general. A bigger LLM that is better at coding is also better at ideation & ethical advice & medicine & math. This isn’t true of everything, jaggedness again (see fiction writing!), but it is remarkably true.

my agents do 95% of the job flawlessly and then freeze at the one button that actually commits. submit the application, publish the post, ship the deploy, all gated behind me tapping yes. trusted to do the work, just not to decide the work is good enough to send