Paridhi Agarwal retweeted
the whole @every frontier team will be re-testing Astra 6.2 and Sol 6.2 LIVE today at 3 PM ET / 1 PM PT!!
@thsottiaux we'd love for you to come on the stream to help us see what we're missing
our mission is to make sure everyone knows where each new model shines (and where it doesn't) so they can use it to do great work. this way, everyone will be able to see how our process works, and we'll be able to show any changes or updates we find!
One of the funny reasons for the difference in vibes between the influencers and the crowd today on 6.2 Sol/Astra is that we gave a few folks access a while ago and then made a lot of changes in the last 3-4 days and I'm pretty sure no re-testing was done within those days and the videos were produced last week.
Don't get me wrong, while the models are a hefty jump compared to their GPT predecessors, they are by no means perfect and there are a few things the teams are excited for the models to get better at in future iterations in terms of taste and niceness of use. Love all the feedback and we hear you!
Building new features into your software doesn’t mean that it has to be more expensive. Here is the workflow @paridhitweets used to reduce costs:
1. Explain what you want to improve and give your agent the context it needs to understand the issue.
Start with a specific goal—cost, speed, or repetitive work—and give your agent access to any relevant code, documentation, and tools.
2. Develop possible fixes.
Ask the agent where your software might be doing unnecessary work and have it suggest changes, with explanations for how each potential fix would help.
3. Measure whether a proposed fix helps.
Have the agent run the same tasks before and after one change, then show you the difference in cost or time.
4. Investigate any remaining problems.
Ask the agent to reexamine the code and test results.
Install the Every agent: every.to/agent?utm_source=x&…
Get the prompts for each step: every.to/context-window/buil…
We wanted everyone on our team to use the @Every Agent, which meant keeping token costs low. We’ve felt firsthand how quickly those costs can dampen excitement about using an AI workflow.
We pass all costs on to users with 0 markup, so the more efficient the agent is, the more they get for their money.
Before launch, @paridhitweets wanted to cut our token costs even further:
• First, we tried Sonnet, with complex jobs passed to Opus. Quality and speed suffered, and costs went up because handed-off requests were processed by both models.
• She cleaned up inefficiencies in our codebase.
• We moved from Opus 5 to Opus 5.5. Response quality improved, and token costs dropped noticeably.
• She cut how much the agent has to read, including from connectors like @GitHub and @Linear.
• She added caching, so the agent reuses instructions instead of loading them fresh each time.
Install the Every agent: every.to/agent?utm_source=x&…
Read more: every.to/context-window/buil…
Paridhi Agarwal retweeted
I'm bearish on dots, grokbot, all the personal agent products - I learned my lesson in the openclaw wave. People don't need another pet to look after, they need cattle because that's how the hamburger is made.
Paridhi Agarwal retweeted
Working With Agents in Slack nitter.gq/i/broadcasts/1qGoNYyeY…
Over the past few months i've been having way too much fun building the @every agent - a coworker that lives in your slack and your whole team can delegate complex tasks to it.
wrote a piece on how it works and why we decided to build it using @claudeai managed agents.
every.to/source-code/why-we-…
shared some insights into how we made the @every agent more token efficient! check it out here
every.to/context-window/buil…
we built this product twice. once as a bot per person, then again from scratch as one agent that lives in your whole slack.
7 months, a lot of rewrites, and it's finally out. proud of this one!
BREAKING:
Introducing the @Every Agent—an agentic coworker in Slack built to help you AI-pill your company.
It comes with:
- Everything you'd expect from a premium company agent. You can delegate difficult, long-running—like opening PRs, writing reports from large datasets, and monitoring social or customer feedback—and get high-quality results all from Slack
- Every's best workflows, pre-installed and always updated. Compound Engineering to Compound Writing. We add more all the time.
- Personalized Frontier Alerts. to help you keep up with new models. When we drop a Vibe Check, the Every Agent will read it and let you know if the new model is relevant to your workflow.
- Zero markup on tokens. What we pay is what you pay—as token prices go down, so does your bill.
You can try it today with $15 in free credits per person, up to $500 per workspace.
After that: it comes with your $30/month Every subscription, plus tokens at cost.
Add the Every Agent to Slack: every.to/agent
Why it's hard to AI-pill your coworkers
If you’re the AI-pilled one at work, you know the problem: you discover something incredible in Claude or Codex, but getting your coworkers to try it is another matter.
The problem is that most of your AI use is invisible. You do it alone, in a terminal or a chat window.
Your coworkers see the finished report or the five new ads, but they never see the request that produced them, or the corrections you made along the way. When you describe it afterward, it sounds like magic, but nobody learns a magic trick without seeing how it’s done.
Why a company agent, in Slack, solves this problem
When you use a company agent like the @every agent you go from prompting in private, to prompting in public.
In a Slack channel, a coworker who would never open a terminal scrolls past a thread with your agent interaction and thinks: I could do that too.
This lets you show instead of tell your coworkers about your discoveries. New AI tools and workflows then spread naturally and virally across your organization.
That's why we made the @every agent for ourselves—and why we're releasing it to you today.
Add the Every Agent to Slack: every.to/agent
This is so insanely cool and useful! Also crazy how incredible fable can be with the right steering
We know today's AI is great at making beautiful HTML one-pagers. I wanted to see if it could make a beautiful deck of interconnected HTML pages that feel like an immersive experience, optimized for learning and ingesting information.
Today I'm sharing a skill that lets Fable 5.1 do this. I'm calling it "Explorable Explanation".
It asks the AI to write an explainer that you will want to dive into: multiple short pages, one idea per page, interleaved with small JavaScript games and interactive visuals you play with so the concept gets seared into your brain.
I've been developing it for the last 10 days, because this turned out to be a surprisingly difficult task for even the latest SOTA models.
Fable 5.1 is the only model that has been able to excel here. In the walkthrough below, I show you what Fable created vs what Astra created from the same prompt. Spoiler alert: Astra was disappointing.
The reason it's difficult is that you're asking the AI to do four things at once, at a really high quality bar:
• Have a really good design and UI sense
• Break the lesson into small chunks and plan them progressively like a teacher
• Write engaging bite-sized lessons per page
• Build interesting toy playables that aid in learning
This requires the working context to split one explainer across tens of files, the empathy of a teacher to predict the reader's curiosity gaps, and the creativity to build toy playables that suit the topic at hand.
Here's the walkthrough video and here's the skill:
github.com/nityeshaga/claude…
If you wanna explore an explorable, check out these:
• nityesh.com/navier-stokes: the big announcement from OpenAI yesterday, explained for someone with high-school math
• nityesh.com/javascript-stack…: the JavaScript web stack, for someone who vibes with it but has never had the foundations explained
• nityesh.com/hotwire: Hotwire, Rails' frontend framework, with a working toy Rails app inside so you can watch what flows between server and browser
1
259
Paridhi Agarwal retweeted
BREAKING:
Introducing Thesis, @every’s annual conference dedicated to answering the most important question in AI:
What does great human work look like after automation?
We believe there is a small group of humans who already know the answer to this question because they’re living it every day. But they're scattered across companies and industries, with few opportunities to learn from each other.
That’s why we’re throwing Thesis November 5th, 2026 in Brooklyn New York.
Our first speakers include:
- Ivan Zhao (@ivanhzhao)—Founder and CEO, Notion
- Andrew Ambrosino (@ajambrosino)—Member of technical staff, Codex, OpenAI
- Cat de Jong - Head of applied AI, Anthropic
- Nick Thompson (@nxthompson)—CEO, the Atlantic
- Josh Miller (@joshm)—CEO and cofounder, The Browser Company
- Cristobal Valenzuela (@c_valenzuelab)—Co-CEO and cofounder, Runway
- Lauren Reeder (@laurenmhreeder)—Partner, Sequoia
- Sahil Lavingia (@shl)—Founder, Gumroad
- Riley Brown (@rileybrown)—Cofounder, Vibecode
- Natalie Fratto (@NatalieFratto)—Founder and creator, Charts & Crafts
- Allie Garfinkle (@agarfinks)—Senior writer and editor, Fortune
- Kane Kallaway (@kanekallaway)—Founder, Wavy Labs
- Nat Eliason (@nateliason)—Head of Founders School, Alpha School
- Kate Lee (@katelaurielee)—Editor in chief, Every
- Katie Parrott (@kplikethebird)—Staff writer, Every
- Kieran Klaassen (@kieranklaassen)—General manager of Cora, Every
(With more very special people to announce soon!)
Learn more: every.to/thesis-2027
What to expect
Thesis will feature talks, demonstrations, working sessions, office hours, and small-group conversations with builders, operators, and execs who are already using AI to do incredible human work.
And, of course, you can bring your agent.
Why New York
New York is where the AI wave hits the beach: It's where new model capabilities meet real-world work. Because of this, it’s the best place in the world to see what happens when frontier technology leaves the lab and enters everyday life.
That’s why we’re holding Thesis at Pioneer Works, a cultural center in Red Hook dedicated to blending art, science, music, and technology.
Space is limited, so we’re accepting attendees by application. We’ll also livestream it for free for anyone who can’t attend in person.
You should apply below.
Apply to Thesis:
every.to/thesis-2027
Paridhi Agarwal retweeted
Chad Review: I gave my AI a bullshit detector
AI slop is getting out of hand — and the newest, most capable models like Fable 5 and Opus 5 do it worst.
And the annoying part: they're not just padding fluff anymore. They reach for the clever word, the metaphor, the caveat nobody needed. They're trying to sound smart.
So I built a skill for it. "Chad Review".
You know Chad from the memes? (see image below) He asks the dumb question without a flicker of shame — which turns out to be exactly the personality of a good bullshit detector.
That's what the skill does. It takes whatever AI just produced and runs a debate where a Chad agent reads it cold, knowing only the job the thing is supposed to do, asking the dumbest honest questions he has.
Why is this here? What does this word mean? Why say it the fancy way instead of the short way? Every time he hits jargon or a metaphor, he points at it and asks.
And it worked!
Here's the part that surprised me. The model that's great at overcomplicating things — is also smart enough to make a really good Chad.
Same model, opposite job: asking the dumb, pointed questions instead of showing off.
How it works
It's not just "hey Claude, simplify this." It runs a real debate, and that produces a more thorough review.
• It clones the artifact first. Your original is never touched.
• Chad asks the dumb questions. He meets the thing cold, with only the job to be done as context, and fires pointed questions at every part.
• A defender answers. Another agent — standing in for the AI that made the artifact — takes each question and either defends the sentence or rewrites it plainer. It's responsible for producing the cleaned-up copy.
• A judge accepts the result and hands you a plan — plus a ranked "add-back menu" of anything Chad sanded off that you might want to keep. You stay in the loop; nothing gets applied until you say so.
What Chad says
I had Chad review a landing page for a repo I'm about to open source, a piece of writing and this post itself.
Here are examples of the kind of blunt feedback it gave me:
• "A one-person publishing house" — it's a blog with a newsletter. Why is it called a publishing house? I almost bounced thinking this wasn't for me.
• "The license is the customizability" — why say it this fancy way instead of "it's MIT, change anything"?
• "A gate screens every page" / "the one canonical line the gate added" — the page invented a word and never defined it.
• "The migration was parity-gated" — what does parity-gated mean? I had to read the diagram below to guess.
• Bird's-eye: Mostly lands — the Chad concept is clear and the two examples are the best part — but it's carrying a naming mess, explains the same idea twice, and ends with zero way to try the thing it's selling.
Chad Review is a Claude Code skill. Grab it here and run it on the next thing your model hands you. It's fun to read what it flags!
github.com/nityeshaga/claude…
TIL that anthropic literally published a prompting guide to deal with the jargon that opus 5 produces and that reading the docs would solve a lot of my problems
3
312
been loving the workflows claude code launches when you ask for a code review. it fans out parallel agents that each own one thing: one hunting correctness bugs, one on efficiency, one checking test coverage.
and the agent verifying a finding is never the one that found it. works great with fable!
2
5
470
Paridhi Agarwal retweeted
Meet Spool — the Loom alternative of my dreams that i built with Fable.
I built it on the side, by chatting with my AI agent on Slack. This very demo was recorded in it.
For years I've wanted a dead-simple screen recorder that just does the basic things right and doesn't come with a recurring subscription. Hundreds of them exist. Somehow not one of them worked the way I wanted. It drove me a little crazy.
So, when Fable was released last week, I built it. Here's what it does:
→ Works without signing in. This whole demo was recorded logged-out. Open it, hit record, done.
→ Instant share link. It uploads while you record (using the WebCodecs API built into every browser), so the moment you stop, your shareable link is basically already there.
→ A download button. In front. For everyone. Loom paywalls this. Screen Studio — which I pay for — doesn't even give it to you on a shared link. It's insane. Spool just lets anyone download the video.
→ Reactions, comments, renaming, and trimming right in the browser. Plus a 3-2-1 countdown — both shipped today.
And here's the part that still feels unreal: it costs me essentially $0 to run. Videos live in Cloudflare R2 (~$1 per 100GB), and the Rails app piggybacks on a DigitalOcean machine that already runs all my side projects. Even if hundreds of you start using it tomorrow, it's maybe a couple of dollars a month.
Cost breakdown:
• Domain - $1/year
• Cloudflare - $1/100GB/mo
• DigitalOcean (shared with 3 other side projects) - $24/mo
How it got built: I never touched Claude Code directly. I asked Luo Ji — my and my brother's personal AI agent that lives in our Slack — to study how we were using WebCodecs in another project, and to build me a Loom alternative with it. 92 messages of planning, course-correcting, and iterating later, I had something real deployed!
That's the thing I can't get over. I created the software of my dreams on the side, by chatting with an agent — and I can now offer something genuinely useful to the world for less than the price of a newsletter subscription.
It's wildly cheap to ship real, useful tools now. So I shipped one.
Check out Spool → myspool.xyz
Open-sourcing it very soon. Let me know what breaks. 📷
such a great read on why automation creates more expert work, not less
We’ve automated every single thing we can @every with AI agents.
And yet there’s way more human work to do than ever. We’ve gone from 4 -> 30 human employees since GPT-3.
I wrote a report on the structural reasons: how AI makes expert competence cheap, why that drives up demand for experts, and why the dynamic only intensifies as we approach AGI.
After Automation: every.to/p/after-automation
183
opus 4.7 finally being better than 4.6
vibe check:
Opus 4.7 feels like it's gotten a lot better recently. Both at coding and writing / strategy / deep thinking tasks
a few people inside of @every have noticed independently. give it a shot if you haven't in the last few weeks!
111


















