Big

Good morning

1 prompt (can you call it even a "prompt"?) - asking Claude Code (Opus 4.6 - 200K Context) to review developers.google.com/crawli… and a client's robots.txt file and compare if our client meeting best practices. 30 seconds later: Current session: 100% Used. Weekly: 14%. It was 0 and 0. Both reset today! You've hit your limit · resets 5:30pm :(

First signals from Anthropic 🤓

I'm sitting on a beach in Sri Lanka watching my AI agency run itself. Here's what just happened in 30 minutes while I was in the water: 8 specialized AI agents — orchestrated on Paperclip — audited 15 moving companies in New York. They crawled their sites. Pulled competitor traffic data. Found the exact technical issues killing their rankings. Wrote a dev-ready fix ticket for each one. Then a 9th agent wrote 5 personalized cold emails per company. Email #2 includes the ticket. Something like: "P.S. I pulled one issue ready to send to your dev team right now: 14 pages have the wrong canonical tag. Here's the fix." 15 companies. 75 emails. 15 dev tickets. All done. I didn't write a single word. I didn't send a single email. I picked the niche, pressed launch, and paddled out. This is what a non-human SEO agency looks like in 2026. We're just getting started. 🦞

I built an AI-powered SEO audit pipeline. 8 agents, 80 checks, competitive intelligence across entire markets. It runs, it works, it finds real issues. I cannot figure out how to let someone else use it. I'm not a beginner, I've built internal tools with auth, dashboards, APIs. 20 years in SEO, 2 years vibe-coding with AI. Built and sold an agency. But there's a gap between "works on my server" and "public product with trials, billing, multi-tenant, proper security." What I actually need: → A boilerplate that handles: auth, multi-tenant, billing/trials, API keys, usage limits, role-based access, proper security → A way to wire my Python/Node backend into a frontend people can actually use → Something that doesn't take 3 months to set up before I can ship the actual product I don't need to learn React from scratch. I need the infrastructure layer so I can plug in my domain expertise and ship. Non-dev founders who've shipped public products: Did you use an existing boilerplate? Built your own? Found a technical co-founder? Something else entirely? Genuinely asking. Every product idea I have dies at this exact step. Thanks guys! 🏄‍♂️ Building from Sri Lanka

I built an AI-powered SEO audit pipeline. 8 agents, 80 checks, competitive intelligence across entire markets. It runs, it works, it finds real issues. I cannot figure out how to let someone else use it. I'm not a beginner, I've built internal tools with auth, dashboards, APIs. 20 years in SEO, 2 years vibe-coding with AI. Built and sold an agency. But there's a gap between "works on my server" and "public product with trials, billing, multi-tenant, proper security." What I actually need: → A boilerplate that handles: auth, multi-tenant, billing/trials, API keys, usage limits, role-based access, proper security → A way to wire my Python/Node backend into a frontend people can actually use → Something that doesn't take 3 months to set up before I can ship the actual product I don't need to learn React from scratch. I need the infrastructure layer so I can plug in my domain expertise and ship. Non-dev founders who've shipped public products: Did you use an existing boilerplate? Built your own? Found a technical co-founder? Something else entirely? Genuinely asking. Every product idea I have dies at this exact step. Thanks guys! 🏄‍♂️ Building from Sri Lanka

The most honest benchmark for agentic reliability I've seen in a while. If seven specialized agents all fail gracefully on an empty React shell, that's actually good system design. The real failure mode is agents that hallucinate content that isn't there. What does your pipeline do when the inputs are garbage?

Building from Sri Lanka?

My AI agents audited a website. 7 specialists dispatched. 50 pages crawled. Results came back. The difference between a workflow, and agentic pipeline with a brain. Every single agent said the same thing: "I can't do my job. There's no content." The site was a React SPA. Our Spider crawled it with curl. Got back empty HTML shells. Every page: same title, same canonical, 12 words of content. A div with id="root" and nothing else. The DOM agent couldn't check titles. They were all identical. The Links agent found zero internal links. The graph was empty. The Performance agent saw a 4KB shell loading a 2MB JavaScript bundle. The IL Strategist said "all pages have zero crawlable content." Every agent correctly identified the root cause: no server-side rendering. But none of them could do their actual job. So I asked all 7 agents: "Based on this audit, what tool improvements would help you do your job better?" Spider: "I need Playwright rendering so I can save what the page actually looks like after JavaScript runs." DOM: "Give me both versions. Raw HTML AND rendered HTML. I want to compare them." Performance: "I need CrUX API and PageSpeed Insights API keys for real-world performance data." IL Strategist: "I need a URL validator script and rendered content to find link opportunities." Links: "The link graph was empty because all links are generated by JavaScript. I literally had nothing to analyze." The Lead Auditor collected all the feedback and created a backlog. Prioritized by impact. P0: Spider needs JavaScript rendering. Without it, every CSR site audit is useless. So we planned it. Not a quick hack. A proper engineering spec: Pre-crawl: sample 10 pages. Fetch each with curl AND Playwright. Compare 6 signals: canonical, title, word count, H1, meta description, paragraph count. If static matches rendered, skip Playwright for the full crawl. If they don't match, trigger full JS rendering. Three modes: SSR (curl only, fast), CSR (Playwright, thorough), Hybrid (smart mix). The Lead decides which mode based on Spider's pre-check. Every sub-issue gets the mode. Every finding gets tagged: found in static HTML, found after rendering, or found in both. The agents didn't just find the problem. They told us what to build. They prioritized it. They wrote the requirements. The engineering loop runs itself. This is what "self-improving agents" actually means. Not marketing fluff. Real agents hitting real walls and creating real engineering tickets to tear those walls down. Next: build the rendering detector, update the crawler, re-run the audit on the same site. Compare before and after. See what the agents find when they can actually see the content. I'm going to surf. The agents will plan, build, implement, and test.

My agent told me “Go surf. I’ll watch the audit.” @dotta

Totally surprised with the findings of our technical auditing agents ❤️🐳 After 10 audits it decided on itself that it will create a set of checks based on the framework/CMS. So now, if website is running on Wordpress it will create specific checks. Amazing

Ok. The agency is running. 😄 The technical SEO department is stable and auditing real sites. Here's how it works: One task goes to the Lead Auditor: "Audit domain .com ". The Lead doesn't do the work. It delegates. First it dispatches the Spider to crawl 50 pages. Spider comes back: "871 pages discovered, 50 sampled, text extracted, link graph built. 16 seconds." Then the Lead dispatches 5 specialist agents in parallel. HTTP, DOM, Performance, Links, Browser. Each runs its own checks against the crawl data. Every finding comes back to the Lead for quality control. False positives get removed. Severity inflation gets downgraded. Only verified findings make it into the report. 50 pages. 81 checks. Dev-ready tickets. A grade. No human touched the audit. The Lead managed the whole team through Paperclip issues. Created sub-tasks, assigned agents, reviewed output. Week 1 of the non-human SEO agency. The pipeline works. Now scaling it. Running on Paperclip by @dotta and @openclaw with @claudeai