Salvaggio: 'Useful' Agentic AI Deepens, Not Resolves, LLM Risk

There's a move that has become common in discussions of AI: pointing to widespread adoption as a rebuttal to criticism. If millions of people are using it, the argument goes, the concerns must be overblown. Writing in Tech Policy Press, Eryk Salvaggio argues the opposite. That people are using language models, he writes, "doesn't make criticism of them irrelevant. It makes it urgent."

The piece centers on agentic AI: systems that, in Salvaggio's framing, "plan," generating code that writes more code and executing multi-step actions across apps and models. He builds on the foundational 2021 paper in which Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Margaret Mitchell described LLMs as "stochastic parrots -- systems that reproduce statistically likely patterns from training data." Agentic systems, Salvaggio writes, "stack these parrots into interacting outputs -- a stochastic flock." The underlying limitations don't disappear when you chain the models together; they compound, and they become harder to see.

One named consequence is "slopware": AI-generated software applications produced faster than they can be meaningfully reviewed, often aimed at short-term problems. Salvaggio's example is precise: the hard-coded variable that makes a single man's household finance calculator work but leads to overdraft fees when used by a single mom. The errors aren't random noise; they're invisible assumptions that cause harm at scale. He also offers a government-facing version: a rural town uses code to write a rubbish-pickup scheduler, only to find it is sending requests to an Excel file that it deletes nightly. And hallucinations, he notes, are "mathematically impossible to eliminate," which means every agentic chain inherits that uncertainty and propagates it through each subsequent step.

The stakes climb further when the deployer is a government. Salvaggio points to pressure on government agencies to use these systems for automating benefits decisions, contract analysis, and regulatory review -- areas where cascading failures can have serious human consequences. Code in these contexts "must be considered untrustworthy until it's verified," and the defining feature of the agentic model is producing output faster than verification can keep pace. Agentic systems also "loop repeatedly, consuming far more resources than more purposefully built software," a cost that concentrates in computational infrastructure. The broader pattern he names is computational solutionism: access to code generation pulls organizations toward solving policy challenges with new lines of code, even when the problem is not a coding problem.

What the piece does not offer is a specific governance framework or a concrete proposal for what meaningful review would look like at the pace agentic systems generate code. The call is diagnostic rather than prescriptive: examine what usefulness means, for whom, and under what conditions. That framing is genuinely useful for procurement and policy decisions, and the question of who gets to answer it is the one the article leaves open.