À PROPOS DE CET ÉPISODE
Most AI agents today run on the same pattern: give the model a goal, dump in a giant system prompt, let it loop through tool calls until it decides it's done. It's flexible. It's also burning enterprise budgets for no good reason.
In this episode, we break down why the generalized agent loop, the default architecture behind most agentic AI, is quietly a tax on an enterprise's margin. Anthropic's own research shows agents can burn 4x the tokens of a normal chat, multi-agent systems up to 15x, and paying a frontier model to rediscover business rules on every single turn adds up fast.
We dig into the false choice everyone gets stuck on: rigid workflows vs. fully autonomous agents. There's a third option, bounded planning compiled into governed execution, where the model plans and reasons but the system decides what's actually allowed to happen. We cover what that looks like in practice (including a case where model spend dropped by 97%), why security should live in the execution graph instead of a stack of prompt instructions, and why bolting hooks onto a generalized loop isn't the fix people think it is.
If you're building or buying agentic AI systems and want to know where the real costs are hiding, this one's for you.
Read the full piece: The Agent Loop Is a Recursive Tax
