Everything Is an Agent Now, and they are arguing with me

In July 2026, the author joined UiPath after working as a Developer Advocate at AWS for over seven years. Upon joining, they discovered that UiPath's agentic stack had introduced non-determinism into their workflows. The author had previously built systems that always produced the same output given the same input, but now they encountered steps with opinions, which were non-deterministic.

The author realized that everything in UiPath's agentic stack could be categorized into two types of steps: those that required judgement and those that could be coded. Deterministic steps were low-stakes and could be handled by coded functions written in Python. These functions had typed input and output, and used the @traced decorator to appear in the platform's trace view alongside LLM calls.

Non-deterministic steps required human judgement and could not be automated. These steps were high-stakes and needed to be approved by a human to ensure correct decisions were made. The author emphasized that the human role in an enterprise setting did not mean no humans were involved; instead, humans moved from doing the work to approving the interesting 4% of it, where incorrect decisions had high-stakes impacts.

The author concluded that the bottom two layers of UiPath's architecture, consisting of coded functions and RPA processes, formed the moat of UiPath's platform. The agentic layer, which introduced non-determinism, was what made UiPath's platform addressable by something other than a human clicking. They noted that the most valuable aspect of UiPath's agentic stack was not the LLMs or agents themselves but the twenty years of connectors and integrations that made it possible for the platform to function effectively.