The short answer
RPA is rules: recorded, deterministic actions replayed against interfaces, unbeatable for stable repetitive tasks and brittle at the first surprise. Agents are reading: judgment applied to variable inputs, tools used toward a goal, at the price of needing evals and boundaries. Neither replaces the other; mature operations script the deterministic majority and give the judgment steps to agents.
RPA earned its decade honestly: millions of hours of copy-paste work automated by replaying what a human clicked. It also earned its reputation for 2am breakage. Understanding exactly why it breaks is the cleanest way to understand what agents actually add.
Two different machines
RPA (robotic process automation) records a procedure and replays it: open the portal, click the third tab, copy the field, paste it there. It is deterministic by design, which is its power (exact, auditable, cheap per run) and its fragility: the bot does not know what a field means, so a moved button or a reworded email defeats it silently. An agent works from the other end: it reads context, decides, and acts through tools toward a goal, which handles variation and costs you the engineering that keeps judgment safe: boundaries, evals, monitoring.
The comparison
| RPA | AI agent | |
|---|---|---|
| Core operation | Replay recorded actions | Read, decide, act through tools |
| Input tolerance | Exact formats only | Variation and unstructured inputs |
| Failure style | Breaks loudly or silently on change | Degrades plausibly: needs evals to catch |
| Cost shape | Cheap per run, expensive to maintain | Costs per token, needs engineering wrapper |
| Auditability | Perfect: every step scripted | Requires design: logs, evals, boundaries |
| Right task | Stable, deterministic, high-volume | Judgment steps: triage, extraction, drafting |
What the breakage reveals
The diagnostic gift of RPA is that its failures are informative. A bot that runs for years untouched is telling you the task was truly deterministic: leave it alone, it owes nobody an upgrade. A bot that breaks weekly is telling you the task was never deterministic: its inputs vary, its interfaces drift, and a human has been quietly feeding it pre-sorted work. That second bot is not an automation; it is a judgment task wearing a script, and the honest fixes are an API integration where one exists, an agent for the reading step, or graduation to owned software when the workaround web has become the system.
A bot that breaks weekly is a judgment task wearing a script.
How they compose
The mature pattern is unglamorous and it works: rules for the rules, reading for the reading. The hierarchy of fixes applies verbatim: deterministic steps go to deterministic tools (scripts, integrations, RPA where interfaces cannot be avoided), and the steps that need a reader (classify this email, extract these terms, decide the routing) go to an agent with evals proving its behavior. Beware the label inflation running the other way: Gartner counted roughly 130 genuine agentic vendors among thousands, and a meaningful share of the costumes are RPA suites with a new adjective. The test is unchanged from the taxonomy: ask what the system does with an input it has never seen.