Get the next conversation in your inbox
One signal-dense note when a new episode lands. No noise.
Value-First AI Daily - Jul 27, 2026
July 27, 2026
Chris moved his operating org to Opus 5 over the weekend and found an instruction he never wrote sitting inside the model itself, telling it not to convene his team. Nico went the opposite direction and hired a specialist 3D crew to build a browser bakery and a city game with Claude and Blender, narrating everything that still does not work.
Key takeaways
On a model migration, diff the behavior, not the benchmark. Chris only found that his roster had stopped convening specialists because he noticed the behavior change and went hunting for the source, which was not in any file he owned and could not be switched off.
Explicit instruction in your own context can override base-model behavior you do not want, but only after you know the behavior exists. That is the real argument against flipping a model over wholesale and assuming everything downstream still holds.
The lift you get from a bigger model is roughly inversely proportional to how much structure you already had. A one-shot user feels an enormous jump; an operator with architected context feels a smaller one and pays far less for it.
Hire the specific specialist instead of hoping the generalist arrives. Nico's 3D quality stepped up visibly at each move from base model to modeling skills to a dedicated crew, and Chris admitted on air he never made that move despite having the roster to do it.
The missing piece in agent-built visual work is the feedback loop. Nico named it exactly: let the builders live-review their own output with screenshots, or they are working blind no matter how good the crew is.
Asking a model to correct a bias can install more of it. The fix for HubSpot bias added five new mentions of HubSpot to the clause; naming the thing you want avoided is priming, and deletion is usually the actual correction.
For agents with network access, the isolation boundary is the control you own. That is the actionable half of the sandbox-escape story, and it does not depend on any vendor publishing anything.
Keep learning
You just watched the work. Now build it.
The next step is doing it yourself — live, with people on the same path.
Get the next conversation in your inbox
One signal-dense note when a new episode lands. No noise.