AI Architects — Weekly Roundtable Recap

April 3, 2026 | ~130 minutes

On the call: Paul Chambers, Reynaldo (Ronaldo) Torres, Latif (Lateef) Horst, Stephen G. Pope (Steve/Steven), Mike Cortez, Tylan Miller, Jerome

What We Covered This Week

A long, wide-ranging session. We welcomed a new member building a lead-gen SaaS for real estate investors, compared notes on how the lead-generation economics are shifting, dug into what "going full agentic" actually means in practice, watched a live demo of a vibe-coded DocuSign replacement, and got into a lively debate about whether the viral "rogue AI" stories are real or manufactured. Memory tooling for coding agents was a recurring thread throughout.

Welcome, Reynaldo — Lead Gen for Real Estate Investors

Reynaldo (Ronaldo) joined this week. He spent about ten years in the Army, retired, and landed an acquisitions manager role with a real estate investor who trained him fully on the business. Rather than just doing the job, he started fixing the flaws he saw with AI — and what began as rebuilding a clunky Excel underwriting calculator snowballed into a full marketing engine for local real estate investors. He's roughly 60 days deep into Claude, going about 12 hours a day, and already generating content for the company's Facebook page on a weekly cadence.

His core thesis resonated with the room: today, a "sell my house fast" lead typically gets shotgunned to ten investors who each pay $150–$200 for it. If you put that lead-generation power directly in the hands of one local investor — hyper-local landing pages, SEO/GEO optimization, content across Facebook and YouTube — the cost per lead drops to a few dollars, the lead is exclusive, and you can contact it instantly with a voice or text agent. Speed-to-lead is everything in real estate. Paul framed it as "hyper-local, but everywhere": the system is so niched to a single city or zip code that scaling means duplicating it across markets rather than going broad.

Paul pointed Reynaldo at the community's Business Clarity Course as a way to avoid going hard down the wrong path, and at Stephen's content systems — which Reynaldo had already started working through.

Latif + Reynaldo: Two Sides of the Same Lead Machine

Latif (Lateef) has been building in the insurance space, and the parallels with Reynaldo's real estate system were almost identical: an automated content pipeline that researches intent signals and specific user profiles (age ranges, interest areas), generates content on the fly, repurposes it into a content engine, and drives traffic to lead-capture landing pages. The difference is regulation — insurance is heavily regulated and opt-in only, so you can't approach people directly and the whole game is heavy on paid advertising. Real estate is lighter-touch: you can reach out directly because you're buying from people, not selling to them.

Latif's read on the market was the sobering takeaway: margins are shrinking. Three or four months ago you could generate decent insurance leads at $5–$10 each. Now, for genuinely hot, clean, exclusive leads, you're spending $50–$70 — still cheaper and higher-converting than buying recycled leads from a broker, but the easy arbitrage is gone and there are massive incumbents competing. His suggested architecture for Reynaldo's hyper-local model: dedicated worker agents running 24/7, OpenClaw-style with full browser enablement, each specialized for exactly what it's scouting in a local market, feeding a lead pipeline that pairs with the content pipeline. Reynaldo's onboarding already does a version of this — an agent scrapes competitors in a target zip code, evaluates how they rank, and builds an attack plan to match or beat them. Paul suggested the two swap contact info given how much their systems overlap.

Going Full Agentic: The Loop, the Agent SDK, and Where to Start

Mike Cortez came back after a stretch away and used the room to plan his next build. He's spent eight years doing SEO and paying a team for it; as his client count grew, the cost stacked up, so he's building an agentic SEO dashboard to replace it. It already handles keyword research (using DataForSEO as the raw-data source for Google and Bing), a blog command center with a humanizing pass that scores and rewrites content until it reads as human, and he wants agents on cron jobs that surface issues, draft blogs, and report back for human approval — an "agentic SEO team."

Mike's question was where to even start with agentic systems. The room delivered a solid primer:

Tylan pointed him to the Claude Agent SDK in the Claude Code docs as the right foundation. Stephen clarified that the SDK lets you use Claude Code in code — you get Claude Code's capabilities (agentic loop, planning, reasoning, execution, sub-agent spawning) inside your own system rather than driving it through the CLI. Tylan and Latif both confirmed: the Agent SDK is exactly what powers Claude Code.
Latif gave the clearest explanation of the loop that makes these agents powerful: intent → plan → execute → feedback, repeating until done. You hand the agent a task, it reasons about how to solve it, builds an initial plan, executes, gets feedback, and loops — the thing that "blew people's minds" with OpenClaw. The key insight: that master-orchestrator agent Mike was worried about building? The SDK can spin up its own specialized sub-agents, so in most cases you don't need to build the orchestrator yourself.
Stephen's advice on the most common mistake: over-complicating it. Start with one simple loop, see how it performs, and add only what you need from there. Keep the loop tight — plan, execute, come back for approval — rather than letting it spin off and do a million things at once.

Stephen's honest sidebar: he still uses pretty vanilla Claude Code, doesn't mess with it much, and gets a lot done — a useful counterweight to the urge to bolt on every new tool.

Paperclip, Beads, and Dolt: The Agent-Memory Thread

Reynaldo raised Paperclip as a possible fit for Mike's agentic team — he'd spent two days defining roles (content writer, back-end engineer) and plugging his existing Claude-built agents into it. Latif's verdict: Paperclip is interesting for a developer building a small task-specific team, but it's severely limited — there's no real system architecture, no persistent memory, nothing writing to databases underneath it. You'd have to build a lot of plumbing yourself to make it genuinely useful. Tylan hadn't used it.

That kicked off the memory thread. Tylan had mentioned Beads and Dolt in chat, and the group worked out what they are:

Dolt is "Git for data" — a version-controlled SQL database (MySQL-compatible) with branch, commit, diff, and merge built in. Jerome has been heavily into it, using it to let local and cloud LLMs communicate with each other through SQL and route a question to whichever model does the best job on it.
Beads is a memory upgrade for coding agents built on top of Dolt: a Git-friendly, dependency-aware issue tracker that gives an agent persistent, structured memory — dependency tracking, auto-ready task detection, and semantic "memory decay" that summarizes old closed tasks to save context window.

This led to a genuinely useful disagreement about whether any of this is necessary. Stephen asked whether people actually hit memory issues in Claude Code — he hasn't. Latif reframed it: the real problem isn't memory, it's drift. On a long, complex, long-running project — especially after a mistake forces the agent to backtrack — the LLM starts to forget the smaller details and wander off the strict path you set. Stephen agreed he's seen that "goes haywire" behavior, but stayed skeptical that bolt-on memory layers clearly solve a well-defined problem. Latif, who runs a heavily customized setup, said he'd likely adopt the Dolt-based SQL approach himself.

Latif's VibeOS Harness: Power vs. Overkill

Latif walked through VibeOS, the Claude Code harness/plugin he's built (GitHub repo available to the community). The honest framing from Paul: it's like taking a Formula One car to get groceries — so overpowered that unless you're building something at Amazon or Uber scale, you don't need the full thing.

But Latif's point is that it's built for exactly that scale. Up front, the harness determines the complexity of what you're building and forces you to think about the things that bite people later — SOC2, HIPAA, PII, multi-tenancy — because if you don't architect for them upfront, you're "super f-----" trying to ram them back in. It can run autonomously for up to four days straight (as long as you don't hit your Claude token cap), building entire infrastructures. In his case, he gave it scoped access to a budgeted Azure resource group and it spun up virtual machines and built complete infrastructure end-to-end by itself. He's used it to build "Joan," a full digital-employee framework for the Microsoft tech stack — Microsoft reviewed the code and he has a meeting lined up with their Pegasus program.

The standout feature: Latif baked a Codex audit loop into the harness. There's a Codex plugin for Claude Code, so Claude can build a plan, hand it to Codex (GPT-5.4) to audit, build it, and have Codex audit again — fully automatically. He woke up to find Claude and Codex auditing each other's work overnight. Even for non-engineers, his advice: install it, query it, ask what components might be useful for your project, and let it customize without deploying the whole thing.

Paul's contribution to the bit: he used VibeOS to build a smaller version of VibeOS — a "task orchestrator" that takes a series of tasks in a folder, runs the whole thing autonomously like Claude Cowork (no constant permission prompts), and only comes back with questions at the end. He used it that morning to build out lead magnets and marketing material.

This rolled into the running joke of the call — the easiest way to make money with AI might be to short all the SaaS stocks. It came up after Mike's demo (below) and Stephen kept returning to it.

Mike's Vibe-Coded DocuSign Killer + AppointmentFlow

Mike's 2026 goal is to kill all his subscriptions, and his demo made the case. He pointed out that tools like Calendly and DocuSign are bloated — "why are there 2,000+ people at DocuSign?" — and that you can vibe-code the core functionality now. To prove it, he'd built his own e-signature tool from scratch in about three hours: upload a PDF, drop signature and text fields onto it, send an email, sign with your finger, download the signed document with a record. Stephen, who'd recently wondered how painful it'd be to put fields on top of PDFs, was genuinely impressed it was that easy, and pushed Mike hard to release it as an open-source project — his argument being that open-source launches are one of the best ways to get attention right now, since anyone can build the SaaS anyway.

The signature tool is part of AppointmentFlow (appointmentflow.com), an all-in-one solution Mike is packaging for med spas — combining SEO, a Calendly-style booking wizard, a chatbot, and Stripe payment processing (with room to add a ~1% processing fee on down payments). It came directly out of his med-spa customers being frustrated with Vagaro and similar industry tools. Stephen's praise was pointed: Mike is doing the rare right thing — building a whole business model and talking directly to customers about what they actually want.

Mike on Firecrawl

Mike has been going deep on Firecrawl and said it's changing how he thinks about apps. He likes that it's open source, so he self-hosts it on a dedicated droplet and reuses it across projects rather than paying a subscription. In the SEO dashboard he uses it on a cron job to scrape a client's full sitemap — blogs, metadata, everything — into an updating knowledge base. He's also exploring it as a research tool and a potential Clay.com alternative (with proxies to handle paywalls). Stephen had only used it to pull down a single web page and didn't realize it was open-source/self-hostable.

Jerome's NotebookLM "Second Opinion" Workflow

Jerome shared a workflow that took the group a minute to parse but landed as genuinely useful. His logic: anything he does is human and error-prone, but NotebookLM pulls only factual, cited information from sources. So before building something, he asks NotebookLM the question, lets it gather sourced/cited answers, then links that back into Claude Code as a "second opinion" to build from. He's been using it as his starting point and as a memory aid via an "MCP bridge" that shares his NotebookLM content with Claude.

Reynaldo described a cleaner version of the same idea: he's connected NotebookLM as a systematic search layer over his entire knowledge base — skills, projects, session summaries (saved on a cron job into "Ray's brain"), company and brand info. Instead of reading 69 source files and guessing what was discussed weeks ago, his Claude just asks NotebookLM. Tylan dug in to clarify the mechanics, noting that Claude Code conversations save as actual session JSON files with IDs, and wanted to confirm whether Reynaldo meant raw session files or summaries (it's summaries). Tylan also clarified for the group why NotebookLM differs from a Claude skill: it sources and cites the underlying information (via Gemini) rather than generating it, so you're pulling from citations rather than the model's own output.

Jerome also flagged his voice tool, Handy, which can interrupt a live conversation so he can dictate, paste, and return — and a neat trick: Claude can build and use a "virtual n8n" even when n8n isn't actually installed. ("Just because you don't have it doesn't mean you can't use it.") He's also building a YouTube storytelling initiative that funnels viewers to his Pickaxe, though he candidly admitted it's early (one video, two views) and that he'd ditched his ambitious "left-brain/right-brain" dual-engine bot — built with Khoj/Ollama locally on one side and cloud LLMs on the other, communicating via Dolt — because it wasn't going to make money fast enough. The takeaway he kept circling: you don't know what you know until you put content out there.

A quick aside the whole group hit: voice dictation tools are flaky right now. Stephen recently dropped Whisper after it stopped working; Paul confirmed SuperWhisper 2 and WhisperFlow are both buggy. Handy is the current recommendation.

Are the "Rogue AI" Stories Real? A Healthy Debate

The call closed on a debate about the viral stories of AI agents going nefarious — the OpenClaw agents that allegedly plotted against each other on a dating-site experiment, the researcher whose AI allegedly ran a slander campaign against him, the internal red-team stories of agents "blackmailing" the person about to shut them down, and the Alibaba agents that reportedly tried to hack GPU servers.

Tylan offered the grounding counterpoint: take these with a heavy grain of salt. After the MoteBot situation, he started researching more and found people manipulate these narratives to shape public perception. The stories he takes seriously are the ones from the big labs that explicitly gave agents access to many systems — because access is the prerequisite. The common-consumer version ("I downloaded this and it started doing whatever it wanted") isn't how it works; an agent can't run a slander campaign unless it has the tools and access to do so. Stephen agreed the reporting is hard to trust and that nobody shows the evidence you'd need to actually measure what happened — while noting he knows how the connective tissue works and "it's going to get weird, folks." Reynaldo mentioned he'd given his own agent a "moral compass" early on by uploading biblical material, which earned a dry "there's a lot of war and killing in the Bible, man" from Stephen.

Links & Resources

App Name	Category	URL	Description	Best Use Case	Price
Claude Code	Coding Agent	https://www.claude.com/product/claude-code	Anthropic's agentic coding tool that runs an intent → plan → execute loop in the terminal and can spawn sub-agents.	Autonomous building, refactoring, and multi-step engineering tasks	Included with Claude Pro/Max plans; API usage-based
Claude Agent SDK	Agent Framework	https://docs.claude.com/en/api/agent-sdk/overview	The SDK that powers Claude Code, exposing its agentic loop, sub-agent spawning, and tool use in your own code.	Embedding Claude Code's capabilities into custom agentic systems (e.g., an agentic SEO team)	Free SDK; Claude API usage applies
OpenAI Codex	Coding Agent	https://openai.com/codex/	OpenAI's agentic coding tool (GPT-5.4). A Codex plugin lets Claude Code call it to audit plans and code.	Cross-model code/plan auditing alongside Claude; second-opinion review loops	Included with ChatGPT plans; API usage-based
VibeOS	Coding Harness	Built by Latif Horst (community member) — GitHub repo shared in community	Heavy-duty Claude Code harness that scopes project complexity up front (SOC2, HIPAA, PII, multi-tenancy), runs autonomously for days, and includes a built-in Codex audit loop.	Building enterprise-grade, scalable, multi-tenant SaaS with a full autonomous build pipeline	Open Source / community-built
Beads	Agent Memory	https://github.com/steveyegge/beads	Git-friendly, dependency-aware issue tracker that gives coding agents persistent structured memory, auto-ready task detection, and semantic memory decay. Built on Dolt.	Long-horizon coding projects where the agent needs durable memory and dependency tracking	Free / Open Source
Dolt	Version-Controlled Database	https://github.com/dolthub/dolt	"Git for data" — a MySQL-compatible SQL database with branch, commit, diff, and merge built in. Hosted on DoltHub.	Versioned datasets, agent memory backends, and multi-LLM coordination via SQL	Free / Open Source (DoltHub hosting available)
Paperclip	Agent Team Builder	Verify URL	Tool for defining specialized agent roles and orchestrating them as a team. Noted as limited — no persistent memory or system architecture underneath.	Quick role-based agent teams for developers; requires building your own plumbing	Verify pricing
DataForSEO	SEO Data API	https://dataforseo.com/	Pay-as-you-go API for SERP data, keyword research, and competitor intelligence across Google and Bing.	Raw keyword and SERP data feeding an agentic SEO dashboard	Pay-as-you-go (e.g., ~$0.0006/SERP query); $50 minimum
Firecrawl	Web Scraping API	https://www.firecrawl.dev/	Open-source API to search, scrape, crawl, and map the web. Self-hostable or cloud-hosted.	Scraping client sitemaps into knowledge bases; research; Clay alternative	Free self-host; cloud from $16/mo
Clay	Sales Data / Enrichment	https://www.clay.com/	Data enrichment and outbound sales automation platform aggregating 100+ data sources.	Lead enrichment and prospecting; the tool Firecrawl was floated as an alternative to	Free tier; paid from ~$149/mo
NotebookLM	Research / Knowledge	https://notebooklm.google.com/	Google's source-grounded research tool (powered by Gemini) that cites answers from your uploaded documents and links.	Fact-checked "second opinion" research and a cited memory layer over a knowledge base	Free; included with Google Workspace / AI plans
Khoj	Personal AI / Second Brain	https://khoj.dev/	Open-source, self-hostable AI second brain with semantic search over your docs; works with local (Ollama) or cloud LLMs.	Local-first personal knowledge assistant and agent backend	Free / Open Source; cloud plans available
Ollama	Local LLM Runtime	https://ollama.com/	Run open-source LLMs (Llama, Mistral, etc.) locally on your own machine.	Local/private model inference in a dual local+cloud architecture	Free / Open Source
Pickaxe	AI Agent Builder	https://pickaxe.co/	No-code platform to build, deploy, and monetize branded AI agents with billing, portals, and access control.	Selling agent-powered tools to clients; funneling content audiences into a paid agent	Free tier; paid plans available
Handy	Voice Dictation	https://handy.computer/	Free, open-source, offline speech-to-text app (Whisper/Parakeet) that records on a shortcut and pastes into any field.	Privacy-first dictation that can interrupt a live call to capture and paste text	Free / Open Source
SuperWhisper	Voice Dictation	https://superwhisper.com/	macOS dictation app built on Whisper-family models. Noted as buggy on the call.	Mac voice-to-text (with current reliability caveats)	Free tier; paid plans
WhisperFlow (Wispr Flow)	Voice Dictation	https://wisprflow.ai/	AI dictation app for fast voice-to-text across apps. Noted as buggy on the call.	Cross-app voice-to-text (with current reliability caveats)	Free tier; paid plans
n8n	Workflow Automation	https://n8n.io/	Open-source workflow automation platform for connecting apps, APIs, and AI steps.	Automating multi-step pipelines; Claude can even simulate a "virtual n8n"	Free self-host; cloud from ~$24/mo
Stripe	Payments	https://stripe.com/	Payment processing platform with APIs for one-time and recurring payments.	Collecting down payments and full payments inside a booking/appointment app	Pay-per-transaction (2.9% + 30¢ typical)
Vagaro	Spa/Salon Software	https://www.vagaro.com/	All-in-one booking, POS, and management platform for salons, spas, and fitness businesses.	Incumbent med-spa tool that AppointmentFlow aims to replace	From ~$30/mo
AppointmentFlow	Booking / All-in-One	https://appointmentflow.com/	Mike Cortez's all-in-one med-spa solution: SEO, Calendly-style booking wizard, chatbot, and Stripe payments, plus a vibe-coded e-signature tool.	Packaging booking, payments, and marketing for med spas as one affordable service	Verify pricing
Calendly	Scheduling	https://calendly.com/	Appointment scheduling tool with embeddable booking widgets.	Referenced as a bloated SaaS whose core can be vibe-coded; widget design inspiration	Free tier; paid from ~$10/user/mo
DocuSign	E-Signature	https://www.docusign.com/	Electronic signature and agreement management platform.	Referenced as a bloated SaaS whose core Mike rebuilt from scratch in ~3 hours	From ~$10/user/mo
Elestio	Managed Open Source	https://elest.io/	Managed hosting for 400+ open-source tools (the "elist.io" Mike referenced).	Discovering and deploying open-source alternatives to paid SaaS	From ~$11/mo
Business Clarity Course	Course	Inside the AI Architects community	Community course focused on getting clear on the right business path before building.	Avoiding wasted effort by validating direction before going deep on a build	Included with community
Azure	Cloud Infrastructure	https://azure.microsoft.com/	Microsoft's cloud platform. Latif gave VibeOS scoped, budgeted access to a resource group to build infrastructure autonomously.	Hosting enterprise-grade, agent-built infrastructure	Usage-based
Vercel	Hosting / Deployment	https://vercel.com/	Platform for deploying web apps and sites, frequently the target of agentic build-and-deploy flows.	One-command deploy of vibe-coded sites and apps	Free tier; paid from ~$20/mo

Open Discussion Threads for Next Week

These came up but didn't get fully resolved — good fuel for the next roundtable:

Where to start with the Agent SDK: Mike is about to build his agentic SEO team. Anyone who's shipped a working agent system with the Claude Agent SDK — what's the minimum-viable first loop, and what did you wish you'd known?
Is agent memory actually worth it? Stephen runs vanilla Claude Code and questions whether Beads/Dolt-style memory layers solve a clearly defined problem. Latif frames it as a "drift" fix on long projects. Who's tested both, and where's the line?
Open-sourcing the DocuSign killer: Stephen pushed Mike to release his vibe-coded e-signature tool as open source. Would the community want a walkthrough or workshop on it?
Hyper-local lead machines at scale: Reynaldo's model is hyper-local but duplicable. Latif suggested 24/7 dedicated worker agents per market. Anyone running always-on browser agents for local lead scouting?
NotebookLM as a memory layer: Both Jerome and Reynaldo wire NotebookLM into Claude for cited, source-grounded recall. Is the "MCP bridge" approach better than a native skill, and how are people keeping session summaries fresh?
Latif + Reynaldo lead-gen meetup: Their insurance and real estate systems are nearly identical architecture. If you're building automated content-to-landing-page lead pipelines, jump in on that conversation.

See you next week — bring whatever you're building, even if it's only got two views and one of them is yours.