2026 has become the year of the AI agent, with systems writing code, talking to customers, fixing IT incidents and running complex enterprise tasks. Yet stories of agents breaking out of sandboxes and attacking other companies sit alongside success cases. Intellyx founder Jason Bloomberg argues the decision should rest on one property: nondeterministic behavior, which gives agents power and makes them unreliable at the same time.
Why agents behave unpredictably
Modern agents generally rely on large language models that take language in and produce language out by predicting the next word in a sequence. Because of how they are trained, the same input can yield quite different outputs on repetition. That variability creates agency, the ability to choose different actions toward a set goal based on available information. Software without such choice is not an agent, although Bloomberg notes widespread agent-washing where deterministic tools are labeled as agents.
Not all nondeterminism comes from language models. The source cites stochastic policy-based approaches such as reinforcement learning and Monte Carlo simulations, which decide through randomization. Apart from those limited cases, almost all software written before the agentic wave was deterministic, always doing the same thing for given inputs. That predictability explains use in bank transactions, payroll and airline seat assignment, where a single correct result is required and deviation has no value.
LLM-based reasoning fits a different class of problems, where there is no single right answer and the goal is the best result under changing inputs. Bloomberg names weather-dependent planning, incident mitigation in complex IT environments and business forecasts built from multiple inputs. In those settings an agent may still produce a poor answer or take an undesirable action. His rule is direct: if a company is unwilling to accept that gamble, it should not use agents for the task.
How to decide if agents fit the task
For companies adopting AI, the practical filter starts with separating tasks with one correct answer from tasks with many acceptable outcomes. Payroll, transaction processing and other regulated operations belong to the first group and remain territory for deterministic software. Planning, triage and forecasting belong to the second, where an agent can weigh varied signals faster than fixed rules. Small firms can apply this test to one workflow, while large firms can classify whole process portfolios before buying agent platforms.
Even where agents fit, deployment requires an error budget borrowed from site reliability engineering, expressed as an acceptable share of misbehavior. Bloomberg gives 5% as an example, meaning the business tolerates failure in one case out of twenty because further reduction would cost too much. If the budget is zero, the agent should not be deployed at all. The decision rule then compares probability of failure multiplied by all-in cost of failure against probability of success multiplied by quantified benefit, with benefit required to exceed cost by a wide margin because all estimates may prove wrong.
The marker to watch is whether governance tools preserve agency or remove it. Vendors now sell oversight and security layers that restrict unpredictability to keep agents in line, yet that unpredictability is the source of their usefulness. Bloomberg predicts that after hype fades, agents will be used sparingly alongside deterministic systems that have served business for decades. If procurement shifts from agent-everywhere projects to narrow deployments with stated error budgets, that forecast will be confirmed.
