Stop Building Agents. Start Building Judgement Around Them.

Your team is shipping AI agents you do not know about. And they are doing it on public links with no login. This week on AI Operators, Dave, Matt, Richard and returning host Joanna work through what the panel is calling the agent sprawl problem. OpenAI just shipped an internal publishing layer inside Codex that gives ChatGPT Business and Enterprise customers a SharePoint-style way to share micro-tools. The timing is not a coincidence. Employees are already building calculators, process tools and brand assets in Lovable, Replit and Codex, and they are sharing them as unsecured public URLs. The new Codex sites feature is OpenAI's answer to the security hole the panel has been warning about for months. From there the conversation opens up. Matt brings a research note that has been sitting in his head all week: models are no longer the product. The control plane is. The harness, the routing, the judgement layer wrapped around the model is where the value is moving. Replit is shipping a multi-agent business package. Codex is shipping a desktop app and CLI. Cursor is shipping Composer 2.5 at one tenth the price. Hermes is shipping a native Windows installer. Everyone is racing to own the same layer. Then the cost reality. Token burn is forcing a generation of operators off their favourite tools. Dave has moved to Hermes running on Minimax m3 for ten dollars a month. Matt hit his Codex cap and pivoted to Cursor's cheapest plan and is barely spending a tenth of the weekly budget. Richard points out the new risk: AI search answers are starting to carry private advertising, and the line between an honest answer and a paid answer is about to get very thin. The hardware moment is next. NVIDIA's N1X chip is the Windows version of Apple Silicon. A 120-billion-parameter model on a laptop, super battery efficient, Jensen Huang's "AI data centre in the palm of your hand." And then the harder truth: power infrastructure is now the binding constraint. Musk wants to put data centres in space. The panel lands on the line of the night: power is the new oil. The closing arc is the most personal one. Dave walks everyone through his full agent refactor journey. ChatGPT projects. Antigravity skill files. Claude Code and Cowork burning through quota after quota. The full realisation that the agents and the skills had been mashed together. The weekend spent getting Hermes to debug itself and write its own updates so it could use Minimax properly. And the final architecture: a handful of soul files that reason, a set of skills that act, and a marketing agent that knows when to call which one. It is a candid, sometimes messy, account of what it actually takes to build an agent stack that does not collapse under its own cost. Plus: the AI Operators IRL event is locked in for 23 June at the Brewery Club in Rouse Hill, northwest Sydney. Twenty seats already gone, the room holds fifty, and Joanna is bringing a statewide AI education push backed by the WA Chamber of Commerce. Chapters throughout. Watch the full episode. About AI Operators: AI Operators is a weekly live show for operators who are tired of the AI hype cycle. Each week Dave Pengelley, Matt Slager, Richard Webbe and rotating guests unpack what is actually working in production, what is breaking, and what it costs. The format is unscripted, ad-libbed, and grounded in real Australian operator experience. No demos for demos' sake. No hand-waving about "the future." Just the next thing to do, and what to ignore until next quarter. Connect with the panel: Dave Pengelley — linkedin.com/in/davepengelley · syllogism.com.au Richard Webbe — syllogism.com.au Matt Slager — slayteksystems.com Joanna Jones (returning host) — emuappleai.com Register for the IRL event: aioperatorspod.com → IRL → Luma #AIOperators #AgentSprawl #AIHarness #SovereignAI #AIOperatorsIRL