unknown

An experienced engineer who built 12+ AI agent systems explains why current autonomous agent hype is mathematically flawed. Error rates compound exponentially in multi-step workflows (95% per step = 36% success over 20 steps), token costs scale quadratically making conversational agents economically unsustainable, and tool engineering for AI interfaces requires specialized design. Successful production agents work within bounded contexts with human oversight, focusing on AI handling complexity while traditional engineering ensures reliability.

Why I'm Betting Against AI Agents in 2025 (Despite Building Them)

Denis Bolkovskis

Just run more agents bro, they will build anything bro

Just run more agents bro, they will build anything bro

The decision makers who decide to invest in AI agents are Dunning Krugered

This is a very very sober take.
I also work in the industry and while at work i think i am building pretty reliable and useful ai tech, especially because we also have things like human in the loop, i also think that you are not completely correct.
Many ai tools are sold with the “AI Agents” pitch but in the end the customers want stateless functions like you described. Most care about Input -&gt; Output and most dont actually want 20 step conversations. That just takes too much time.

Before running agents, it is the {thinking of work} that is done by agents. Design of labour as assembly line management is a new paradigm.