📝 SUMMARY This week's call packed a lot into two hours: deep dives from Patrick Chouinard on his Beacon/Clara research pipeline, Astra-driven website rebuild, and skill-management philosophy; a business-growth discussion with Juan Torres on his AI photo booth; Rod Morrison's return with questions on operationalizing agent workflows; a live demo of Ty Wells' interactive pitch app for his Bahamas ERP/CRM platform; and an introduction from new member Lisa Jetton. The second half turned into a rich round-robin on personal infrastructure — Proxmox setups, T3 Code environments, multi-orchestrator stacks, and an open-source visual LLM project — plus updates from Ryan C and Morgan. Beyond the specifics, the recurring theme (voiced by Patrick near the end) was that the group itself serves as a weekly "context upgrade": a fast-moving AI landscape is impossible for any one person to track alone, and this community is how we keep up. If you missed it live, the recaps below are worth your time. 💡 KEY INSIGHTS Split your agent stack by concern: Patrick's "Beacon" does raw web research with zero personal context, then hands off to "Clara," which knows everything about him and turns it into a personalized morning briefing delivered via TTS/Discord. Archive everything your agents produce. Patrick keeps full transcripts + audio of every briefing so he can feed prior outputs back in and continuously improve — a self-optimizing loop. Let agents run autonomously with a clear role. Patrick had Astra (acting as "lead of a dev team") dispatch sub-agents and run ~17 hours on a website rebuild, only checking in when needed. Heavy agent building burns tokens fast — Patrick and Ty both blew through $200-tier subscriptions within a week. Budget for it. Turn your existing prompt templates into skills: decompose long templates into an "uber skill" that orchestrates sub-skills, add YAML frontmatter, and define which are callable vs slash commands. Keep a permanent, account-level tech stack definition so Claude and your agents stop recommending inconsistent tools and default to your approved stack.

This is both interesting and scary Dwarkesh Patel has great videos. His guest researcher, Ajeya, fills in the gaps about what happened and provides some impressive insights.

AI Developer Accelerator — Coaching Call - September 15 VIEW RECORDING - 83 mins (No highlights) Meeting Purpose A weekly sync on AI projects, strategies, and development tools. Key Takeaways - Agentic Workflows Drive Efficiency: Custom agents (e.g., Patrick's "Beacon" news aggregator, Ty's "UI in reverse" code generator) are automating complex tasks, but this high-volume usage is rapidly consuming token limits. - Strategic Pivots for Revenue: Juan is shifting from development to sales for his AI photo booth, creating bespoke transformations for high-end venues to secure initial clients. - Operationalizing AI for Consulting: Rod is exploring a consulting product to help SMBs operationalize AI, focusing on creating custom skills and structured workflows to reduce token waste. - New Member Onboarding: Lisa, a full-stack engineer, joined to learn about agentic systems and is evaluating development harnesses like Hermes, T3 Stack, and PI. Topics Agentic Workflows & Token Management - Patrick's "Beacon" News Aggregator: - Workflow: "Beacon" (research agent) → raw report → "Clara" (contextualization agent) → TTS audio → internal archive (audio + transcript). - Goal: Create a continuous improvement loop by feeding user feedback from ChatGPT back into the system to refine future briefings.

Last week we watched a production cost drop from ~$1 per run to ~$0.09 with a single model switch — and somehow that wasn't even the best conversation of the night. Between RAG debugging war stories, a pendant recorder that logs your whole life, and the memorable line "humans are the bottleneck, not AI," there was a lot to chew on. If you missed it, the recap is worth your time — and Tuesday's call is where the follow-ups happen. 📞 HOW THE CALLS WORK The calls can run 2+ hours. We want to make sure we're respecting everyone's time. Especially those of you who actually show up. Here's the structure: 👉 Reply to this post with your questions before the call 👉 If you submit a question and you're on the call, you go first 👉 We work through questions in the order they came in 👉 Then we open it up for everyone else If you can't make the call but want your question answered, drop it in the comments. We'll get to it. But priority goes to people who are there. The goal is simple: if you're taking the time to show up, you shouldn't have to wait behind questions from people who aren't even on the call. Plenty of threads still open from last week: Hemal's RAG model-flipping question is headed to the forum, Tim's security review with Ryan C is in motion, and Ty launches his ERP/business-advisor product on the 15th — the very day of this call. If you've got results, questions, or a "here's what happened next," bring it. And if you've been quietly configuring your AI to be more brutal and less agreeable, we want to hear how that's going. 🔗 ZOOM LINK (save this) https://us06web.zoom.us/j/81995207847?pwd=Xe6u6LmIQOmCP5VTnOwWYjDBfZNKGB.1 📅 WHEN Tuesday September 15th at 6PM ET Looking forward to seeing you on the call!

📝 SUMMARY This week's call packed a lot into one session, moving from practical debugging and cost-saving tactics to career mentorship and, finally, a genuinely thought-provoking discussion about staying active thinkers in an AI-curated world. Members shared hard-won lessons on RAG debugging and model selection, live demos of home-built observability tooling, and real numbers on cutting production costs by switching models. A recurring theme throughout was that technical skill alone isn't enough — whether you're hunting for paid AI work, scaling a business, or getting better answers from your tools, the edge comes from combining technology with domain knowledge, relationships, and deliberate intent. The conversation closed on a memorable framing: with models this capable, humans — not AI — are now the bottleneck on connecting insights, so it's on us to configure our tools to challenge us rather than comfort us. Whether you attended live or not, there's something here worth stealing for your own work. 💡 KEY INSIGHTS Set smart-model baselines first — otherwise you can't tell if failures come from the model, retrieval, or prompt. Check your RAG retrieval before blaming the LLM: junk in, junk out. Long agent sessions re-send the entire growing context (500k–800k+ tokens) each response, burning limits fast — one-shot requests are cheaper than always-on agents. Switching a production workflow fully to GLM 5.3 cut per-run cost from ~$0.80–$1.10 to ~$0.09 with no perceptible quality drop. Use OpenRouter's independent task-based model rankings (coding, agentic, etc.) as a starting point for comparisons. Build observability around your AI usage: track tokens, session status, and unresolved attention across agents — treat LLM usage like infrastructure needing monitoring. GPT's computer-use can functionally replace a QA hire — it finds bugs and proposes fixes for another model to implement. Astra (GPT-6/Codex) beat Fable on a multi-hour admin planning task, and is less restrictive for security audits.