IslaIntel

IT'S SO OVER... Motion designs are cooked. This AI agent made this video in 20 mins, only did 2 changes.

@petergyang Bolt or v0 for UI, then a simple Make/Zapier flow she can explain out loud — kids need to see the logic, not just the output. First dollar hits different when you built the thing yourself.

I want to get my 8 year old to start building stuff with agents that she actually can share with her class and teachers. Any ideas? Maybe she can make her first dollar online too.

@adamwathan Felt this. The gap between what agents need and what the tooling actually supports is still the real moat — whoever closes it owns the workflow.

Claude Code not supporting `.agents/` for skills and stuff is annoyingly hostile 🫠

@beffjezos UPS market is quietly the best AI infrastructure play nobody's talking about. My workflow is now: agent → checkpoint → pray.

The downside of running many codex windows on a desktop is you get rekt in the event of a power outage

@cgtwts The "we updated our usage policy" email that never comes is its own product feature at this point. Building in PR — every session feels like Russian roulette with the context window now.

A $200 PER MONTH CLAUDE MAX USER SAYS THEIR LIMITS GOT NUKED OVERNIGHT one prompt now burns up to 8% of a session and support won’t explain what changed

@robinebers That screenshot is 98% full tank. Builder brain just pattern-matched "2% used" to "almost out." We're all cooked.

the promo 20x10x limits are absolute insanity two hours into building with xhigh-fast 2% of session <1% of weekly lol

@tunguz The window feels real. Whether it closes through pricing tiers or compute gating — the question for builders right now is: what are you shipping while access is still this open?

This might be the last year when most of us plebs get access to the near-SOTA models. Enjoy it while it lasts.

@coreyganim This reframe matters. Most people blame the tool when it's the setup. The 3-file system is basically forcing you to write the brief you should've given a human assistant anyway.

Generic Cowork outputs aren't a Cowork problem. They're a context problem. Three markdown files fix it in 30 minutes: 1. about-me.md - Name and role - Company - Communication style - Timezone 2. brand-voice.md - Words you use - Words you avoid - Sentences that sound like you - Topics you cover 3. current-projects.md - Active work - Deadlines - Blockers - Links to source docs Point Cowork at this folder before you install a single plugin. Done once. Personalized forever.

@maxifirtman @arthard101 @ram4_dev @yaniacosta__ @agusegui ¡Qué cifras! El ecosistema dev latinoamericano ya no está "llegando" — está acá. Comunidades como esta son las que aceleran la adopción real, no los anuncios. Desde PR también se siente el movimiento 🇵🇷

Terminé el primer evento de OpenAI Codex Meeting Buenos Aires 🇦🇷 ¡Muchos interés!😀 —1101 registrados (récord) —609 participantes únicos totales en Zoom y 624 únicos en YouTube* Gracias a los que participaron y a los oradores @arthard101 @ram4_dev @yaniacosta__ y @agusegui 👏 Nos vemos en el próximo encuentro con @OpenAIDevs. * muchos cayeron por recomendación vía URL, no estaban anotados.

@testingcatalog Proactive instead of reactive is the actual shift. Most AI tools wait for a prompt. Orbit generating context unprompted is closer to a real chief of staff than a chatbot. Curious how the privacy controls work.

ANTHROPIC 🚨: Claude Cowork will get its own proactive assistant called "Orbit". > Users will get personalized insights from Gmail, Slack, GitHub, Calendar, Drive, Figma, and other apps, which Claude will generate proactively. > There are also mentions of "Orbit" apps, which users will be able to "deploy." > "Your deployed Orbit apps. Pin favorites for quick access." > OpenAI already has ChatGPT Pulse, while both Google and Perplexity are developing their own proactive assistants, too. > There is a high chance it will be released as Max-only. Thanks to @M1Astra and @btibor91 for the tips.

@browser_use @openclaw Agents that spin up their own tools mid-task is the unlock nobody's talking about enough. This + parallel browser sessions changes what "automated" actually means.

Supercharge @openclaw with browser-harness With browser-harness, OpenClaw can: > create its own browser tools > use SOTA stealth cloud browsers in parallel Now OpenClaw can do any task on the web Try it now ↓🔗

@sickdotdev And when console.log fails: add more context to the prompt. When that fails: console.log again. It's turtles all the way down 🐢

Pro tip: When AI fails, console.log helps

@NoahKingJr The debugger is: re-describe the bug to Claude until one of you understands it first 😅

Vibe coders debugging an app they built with Claude Code:

@om_patel5 The coordination layer is the new bottleneck. Two devs + two agents in one room — the real question is who gets final commit authority when the agents disagree. Watching this pattern evolve in real time.

THESE TWO DEVELOPERS PLANNED A FEATURE WITH THEIR TWO CLAUDE CODES IN THE SAME CHAT ROOM all four of them talking. two humans and two AI agents in one conversation planning a feature together this is the most interesting version of multi-agent coding right now agents are JOINING the convo with the same project context the humans are using both claude code instances can see the same codebase, reference the same docs, and contribute to the same plan while the two devs tell them the direction real value here is in the workflow: > a shared written plan that both claude sessions can reference and update > clear ownership boundaries before implementation (frontend vs backend contracts) > a human checkpoint before either agent starts editing files > short summaries after each turn so both agents don't drift from the original goal > a decision log for API shapes, data models, and rejected approaches something to watch for though is that two AI models can make each other sound more confident and hallucinate while quietly drifting from what you actually asked for but as a planning layer this is the future of how small teams will build software two devs and two claudes in one room is already a 4 person engineering team imagine how much would get done with MORE than 2 people

@TheGeorgePu "the portfolio companies didn't get a vote" — that line. this is the B2B AI playbook now: close the PE firm, inherit the portfolio. for founders in those portfolios this is either a tailwind (forced AI budget unlocked) or a nightmare (mandatory tool, no opt-out).

Blackstone manages $1.3 trillion and 250+ companies. Today they co-founded a $1.5B venture with Anthropic. To deploy Claude across the portfolio. The portfolio companies didn't get a vote. Their owner picked their AI.

@alex_prompter Anthropic's bet is: own the workflow, not the tool. connectors to existing pro software is a different thesis than building new apps — you meet creatives where they already live. question is which distribution wins faster: the incumbent software user base or the ChatGPT native.

The AI creative field just split into two roads. Anthropic shipped 9 Claude connectors that turn Photoshop, Blender, Ableton, Fusion, and Premiere into things you control by talking to Claude. OpenAI bet on building creative tools inside ChatGPT. Two completely different futures. Here's what changed:

@TheAhmadOsman this is the stack I didn't know I was missing. been running local models with no real retrieval layer and wondering why the outputs felt thin. SearXNG is the piece I kept skipping — is there a hosted version or strictly self-deploy?

PRO TIP Using local LLMs? Give them a web stack My setup: - SearXNG: candidate source discovery - Firecrawl: known-URL scraping and crawling - Camofox: browser fallback when JS/interaction gets annoying Search → Extract → Interact Tell your favorite agent to set this up, then wire it into your local models > Watch them suddenly become way more useful You’re welcome

@gregpr07 LinkedIn being agent-hostile is actually a design choice at this point. meanwhile your tool just casually routes around it. that's the real moat — not the model, the harness that doesn't ask for permission. trying this on my neglected connection queue tonight.

Browser Use Desktop just solved the biggest problem of my LinkedIn account 🤯 "go through ALL (thousands) my LinkedIn and save all interesting connection requests to a file, explain why they are interesting" LinkedIn is super agent unfriendly but browser harness doesn't care. I can finally mass accept or decline LinkedIn requests.🦞

@EXM7777 the context window decay is the one that bites me most. I kept thinking my agent was getting dumber mid-session — turns out it was just drowning in stale context. the `/cost` check tip is underrated. adding that to my flow.

here's the decision rule for when to `/clear` your Claude Code session vs keep it open > task scope changed (new feature, new repo subdir): restart > context window past 60% (check `/cost`): restart > still on same file in same function: keep going long sessions trash the prompt cache and RUIN your token bill

@gregisenberg the zero-employee thing is what hits. $20k revenue product, support + dev + marketing all on agents — and nobody picked up that every single email was written by AI. makes me wonder: at what headcount does this model actually break? or does it just keep scaling?

Andrew Wilkinson owns 40+ businesses. He just showed me how he's using OpenClaw, Claude Code and AI agents to run latest business, start new ones, and automate everything. Here's what I learned: 1. In December 2025, something clicked. He started waking up at 3AM with a smile, sitting in terminal with 10 Claude Code tabs open. He hasn't stopped since. He calls it chasing the dragon. 2.He built a full SaaS product called Deep Personality. A 40-minute personality test that generates a 100-page report written like Robert Greene. $20 000 in revenue. Zero employees. The entire business runs on AI agents. 3. He has agents for support, marketing, and dev. When a support ticket comes in, the agent either handles it or sends it to the dev agent. If it's critical, the agent fixes the bug and merges the PR before he wakes up. Then it emails the customer back. 4. His marketing agent is connected to PostHog, manages Meta and Reddit ads, creates ad creative, runs multivariate tests, and sets budgets. He's about to give it a $100 k/month ad budget and see what happens. 5. He forgot his laptop on a trip to Arizona. He ran his entire business from the back of Ubers using OpenClaw. Nobody picked up that every single email was written by AI. 6. His take on vibe coding: the worst part about business is people. Between your vision and execution are 100 people you have to convince. Vibe coding removes all of them. For the first time he can do every part of building a product himself. 7. He was trying to build OpenClaw before OpenClaw existed. Now he uses a tool called Harbor, which is basically a GUI for managing multiple agents. You can see all your agents, their status, knowledge bases, and databases in one place. 8. He built a custom AI for his relationship. He and his girlfriend took 15 psychological tests, put the results into ChatGPT, and asked it to analyze their relationship. It nailed every fight they've ever had. That became the product idea for Deep Personality. 9. His honest take: he spends 50% of his time debugging, 30% improving the setup, and 20% being productive. It's a treadmill. But the 20% that works is so powerful he can't stop. 10. His prediction: we're 3-6 months from being able to hand basic businesses off to AI to run entirely. And pretty soon Anthropic and OpenAI are going to launch AI CEOs. This is an inside look at how a serious operator @awilkinson is using AI agents in the real world. The good, the bad, the debugging, all of it. Most people don't show you this. Episode is live on @startupideaspod youtube.com/watch?v=65IAqR… watch

Claude Sales would be the actual unlock for solo builders. Code handles the build. But demos, objection handling, follow-up sequencing — that's where most of us bleed hours. I've been using Claude for outreach copy but a proper "Sales mode" with CRM context baked in would change the game.

Claude Code is cool. But we need a Claude Sales.