AI Agents, Vibe Coding & Automation

Move from prompts to agent systems; AI is dumb without context; keep a human in the loop.

AI was the connective tissue of the 2026 event. Across very different talks (terminal harness engineering, brand-saturation SEO, local ranking forensics, push-ad newsjacking, agency upsells, and direct-response funnels) speaker after speaker arrived at the same place: the people winning are not the ones with the best one-off prompts, they are the ones who built repeatable systems of AI agents that do the work continuously. The framing shifted from "use AI to write content" to "engineer agents, harnesses, and boards that produce strategy, sites, video, and ad campaigns at scale," paired with a hard insistence that the output is only as good as the context and human oversight you feed it. This brief synthesizes across the vibe-coding session (Andrew of Thorbit.ai; Elias Levadaros), the Kato/Kurtz session, the day-2-part-2 Hartzer/Hawkins/Merlino session, the day-2-part-3 upsell session, and Chris Morrow (day 3).

The through-line

1. AI is "dumb" without deep context; systems exist to engineer that context. Chris Morrow (day 3) made this his mantra ("AI is dumb without context"), arguing any LLM produces generic output until you feed it extreme structured context about the business, customer, offer, and awareness level, and that his nine-to-ten-prompts-in-sequence approach beats a single cheap prompt. Andrew of Thorbit.ai (day 1) reaches the same conclusion from the engineering side with his "cold-start problem": no memory persists between agent runs, so the brief is the only thing the agent knows, and a vague brief is a system-design problem, not a writing problem. Brian Kato (day 2) operationalizes context as MuVERA plus BERT coverage and EAV tables so the agents reason from many angles rather than promotional hype.

2. Move from one-off prompting to repeatable agent systems ("harnesses" / "boards" / "agentic workflows"). Andrew (day 1) names this directly with his three tiers (vibe coding, vibe engineering, harness engineering) and his point that raw vibe coding can produce good output once but cannot be reliably reproduced. Brian Kato (day 2) builds the same idea as a Claude "Board of Directors" of about twelve agent personas that debate, reach consensus, and run cascading parallel waves. Michael Merlino (day 2 part 2) runs a live agentic setup (phonetically "OpenClaw," "Hermes," plus custom bots and a follow-up agent "Hawkeye") monitoring RSS and sitemaps. The unnamed day-2-part-3 presenter has an agent "Oliver" that builds working websites from one prompt. The convergence: stop hand-prompting, stand up agents that run on their own.

3. Adopt agentic AI or fall behind (the survival framing). Merlino (day 2 part 2) is most explicit, warning that agencies failing to adopt AI/agentic workflows will lose to rising churn and falling client budgets, and that the first symptom is heavy payroll. The day-2-part-3 presenter frames AI services as both a revenue and a retention play ("intro the AI shit and what you can do" to keep clients around). Morrow (day 3) cites his friend Tyler's Facebook Ad Library scraper claiming 36 times more advertisers spent at least $1,000 then quit than have currently active ads, reinforcing that most operators give up where automated systems persist.

4. Newsjacking and content at scale is now an agent pipeline, not manual labor. Dan Kurtz (day 2) runs Google Trends CSV exports through Claude agents ("Google Trend Brief Writer") that rank, filter, research, and write advertorial pages with ad slots predefined. Brian Kato (day 2) live-builds keyword and geo-grid spreadsheets (a Boston example of 23 neighborhoods, 80 keywords across 6 clusters) from agent waves. Merlino (day 2 part 2) and Kurtz both mention piping filtered RSS/Feeds into client sites. The shared move: point agents at a trend or geo source and let them produce the strategy artifact.

5. Keep a human in the loop; the agents are wrong sometimes. Brian Kato's (day 2) caveat is blunt: "trust but verify... always have that human in the loop, AI screws up." Andrew (day 1) builds verification into the harness itself via hooks and a "golden master test." Morrow (day 3) frames his whole product around extracting context precisely because the raw model output is unreliable. The automation does not remove judgment; it relocates judgment into the system design and the review gates.

Tactics & playbook

• Name folders after the work (the domain), not the code type. Andrew (day 1): a "estimates" folder for a construction company lets agents grep and navigate intuitively; he claims roughly 50% fewer tokens and AI working about twice as fast.
• Keep the main context window under about 100k tokens; use thin sub-agents to fetch context and report back. Andrew (day 1): he claims performance drops off past ~100k and hallucination climbs after 200k, so deploy many small sub-agents rather than loading one giant window. Check your starting tokens with /context in a fresh Claude Code terminal.
• Use hooks as the anti-hallucination and automation layer. Andrew (day 1): any executable can fire on Claude Code events; a "stop" hook (the "Ralph Wiggum" pattern) sends the agent back to its original spec and loops it; other hooks block actions until a required file is written, or strip stale code comments. Mantra: observability, enforceability, dependability, scalability.
• Build a Board-of-Directors agent project in Claude with consensus and logging. Brian Kato (day 2): upload skill files plus about twelve markdown agent personas (content director, strategic analyst, PM, SEO/PPC specialist, viral video specialist, neuromarketing psychoanalyst, etc.), instruct them to reach consensus, "start every conversation by creating a log.md" so you can reload context after the window fills, and "invoke cascading waves of agents asynchronously in parallel." Output is a strategy deck he sells for $2,500 to $3,000.
• Generate strategy and geo-grid spreadsheets from one prompt. Brian Kato (day 2): prompt the board to "create a marketing strategy for [website]" or set up a geo grid; it produces multi-tab spreadsheets (keyword stack, content stack, search intent, long-tail, competition level, notes).
• Newsjack via a Google Trends to Claude pipeline. Dan Kurtz (day 2): export Trending Now (last 4 hours) as CSV into the brief-writer agent, let it filter unwanted categories and pick the top angle, then hand the brief to a second agent that writes the article and outputs an artifact with ad spots already defined; host on GitHub Pages or Cloudflare Pages on a matching purchased domain. Disable the menu and footer so only your offers are clickable. Swap the prompt to produce local lead-gen pages (a roofer storm-preparedness page with click-to-call or a CRM-wired opt-in).
• Use agents to translate raw analytics into plain client language. Elias Levadaros (day 1) prescribes asking the AI advisor to cross-reference GSC and GA4 and explain a drop in one sentence a client understands in ten seconds (his "so-what gorilla"). Merlino (day 2 part 2) routes missed-call follow-up through his "Hawkeye" agent.
• Let an agent build the website. The day-2-part-3 presenter's agent "Oliver" produced a real, non-WordPress site from a single prompt "while I got dressed." Morrow's (day 3) agentic system built a 6-minute business-acquisition VSL narrated by an AI British voice.
• Encode your peers' expertise into your agents. The day-2-part-3 presenter folds named colleagues' SEO and schema knowledge into his automated processes ("the SEO has Marty Marion's brain in there"). Morrow (day 3) ships "compartmentalized versions of his brain" such as "sales coach Chris."
• Extract business context with a fixed interview, then let the system market continuously. Morrow (day 3): Pocket Marketer's 12-question interview captures context, after which "I will market for you forever." Match copy to one or two of Gene Schwartz's five awareness levels at a time, since AI cannot make copy relevant to all at once.
• Productize a simple AI upsell to existing clients. The day-2-part-3 presenter: introduce AI capabilities to current customers as both a revenue and retention move, and use CallRail to see which platform (Instagram, TikTok) each inbound lead came from.

Tensions & disagreements

• Claude Code (terminal/harness) vs no-code or chat-based building. Andrew (day 1) works ~50 hours a week in the terminal with Claude Code and treats harness engineering as the serious path. Brian Kato (day 2) and the day-2-part-3 presenter ("Oliver") drive agents through the Claude desktop/project UI and skill uploads, not the terminal. Morrow (day 3) loves Lovable ("my muse," used daily) yet says his own funnel-hacker tool "makes Lovable look like a joke for what we do." The referenced (separate) Aaron Gruenke deck explicitly recommends Lovable.dev over Claude Code for non-coders. So the room splits on whether real leverage requires the terminal or whether no-code/chat tooling is enough, and even on which no-code tool wins.

• How much to trust the model's output. Brian Kato (day 2) and Merlino (day 2 part 2) push heavy automation (boards of agents, 250 self-managed GMBs, bot fleets) while Kato simultaneously insists "always have that human in the loop, AI screws up" and Morrow (day 3) argues the model is "dumb" and useless without heavy context engineering. The optimism about volume coexists uneasily with strong warnings that the raw output is unreliable.

• Maximalist context windows vs lean ones. Andrew (day 1) calls huge advertised context windows (he cites a 2M Grok figure) a "money grab" and insists on staying under ~100k tokens. Other speakers who lean on big multi-agent Claude projects (Kato's parallel waves, Merlino's always-on agents) implicitly rely on large-context, always-loaded setups. This is a genuine architectural disagreement about whether to constrain or expand the window.

• Replace humans vs retrain them. Merlino (day 2 part 2) calls simply replacing humans with AI "unethical" and argues for retraining low-level workers even as his own team shrank from 27 to 15. The day-2-part-3 presenter frames AI agents as things that "work for me all the time" with no such caveat. Same automation, different stance on the labor consequences.

• Generic AI content is good enough vs it actively hurts you. This theme overlaps the wider event tension on AI content quality. Brian Kato (day 2) uses MuVERA/BERT specifically to escape "generic" AI output, and Joy Hawkins (day 2 part 2, citing Kyle Roof) notes AI content is bad at putting the exact words you need to rank into the page. The pro-automation speakers assume agent output is publishable; the SEO-forensics voices warn it often is not without engineering.

Sources (conference sessions)

Conference session references, not pages on this site:

• Vibe-coding session (Andrew of Thorbit.ai on harness engineering / vibe coding; Elias Levadaros on the AI-advisor "so-what" translation)
• Kato/Kurtz session (Brian Kato's Claude "Board of Directors" and MuVERA; Dan Kurtz's Google Trends to Claude newsjacking pipeline)
• Day 2 part 2 (Michael Merlino on agentic workflows, agent fleets, and the adopt-or-fall-behind survival warning; Joy Hawkins on AI content limits)
• Day 2 part 3 (unnamed presenter on the AI-services upsell, CallRail lead attribution, and the "Oliver" site-building agent)
• Day 3, Chris Morrow on "AI is dumb without context," 12-question context extraction, and Pocket Marketer

Related Spearleaf system: the entity, distribution, and tiered link-building system that Kato's brand-saturation thesis, Elias's SEO Neo tooling, and Merlino's revenue-over-rankings stance feed into. See DAS campaigns and anchors. Related briefs: Digital PR & conversion and AI search visibility.