How do I tell a real agent from a washed one?

The fastest tell: ask 'if I turn off the LLM, how much still works?' A genuine agent collapses—the model is its planner. A washed one keeps running its script and only loses the wordsmithing. Other tells: you can draw its full control flow as a flowchart, it has zero or one tools, it can't remember prior steps, and it has no failure story beyond 'it escalates to a human.' Score every product on the Levels of Autonomy (AI Agents) rubric—the honest floor for calling something an 'agent' is L2.

Is agent washing always dishonest?

Not always. Many products are genuinely on a path toward autonomy—a copilot today, a semi-autonomous agent next year—and vendors describe the destination in the present tense. The gap can be a hope rather than a lie. Either way you inherit it, so evaluate the artifact in front of you today, not the roadmap around it.

Agent Washing

Written by Max Zeshut

Founder at Agentmelt · Last updated Aug 19, 2026

The practice of rebranding a product as an 'AI agent' or 'agentic AI' without adding the capabilities the term implies—so chatbots, Robotic Process Automation (RPA) scripts, virtual assistants, and rules engines get sold at the agentic premium. Coined by Gartner analysts, agent washing is the 2026 buyer's central due-diligence risk: the defining property of a real AI Agent is that it decides its own steps *at runtime* over a goal, while a washed 'agent' follows a path a designer scripted *at design time*. Because 'agent' has no enforced definition, every vendor draws the line just below their own product. Gartner estimated only a small fraction of the thousands of vendors claiming agentic AI ship anything genuinely agentic, and predicted over 40% of agentic AI projects will be canceled by end of 2027—largely because buyers bought the label, not the capability.

Example

A vendor sells a 'fully autonomous support agent.' On inspection it's a decision-tree chatbot: you can flowchart every path, it has no tools beyond drafting replies, it keeps no state across steps, and it stalls the moment a request leaves the scripted happy path. That gap between the label (Autonomous Agent) and the artifact (an L1 assistant) is agent washing.

Frequently asked questions

How do I tell a real agent from a washed one?: The fastest tell: ask 'if I turn off the LLM, how much still works?' A genuine agent collapses—the model is its planner. A washed one keeps running its script and only loses the wordsmithing. Other tells: you can draw its full control flow as a flowchart, it has zero or one tools, it can't remember prior steps, and it has no failure story beyond 'it escalates to a human.' Score every product on the Levels of Autonomy (AI Agents) rubric—the honest floor for calling something an 'agent' is L2.
Is agent washing always dishonest?: Not always. Many products are genuinely on a path toward autonomy—a copilot today, a semi-autonomous agent next year—and vendors describe the destination in the present tense. The gap can be a hope rather than a lie. Either way you inherit it, so evaluate the artifact in front of you today, not the roadmap around it.

Related glossary terms

Related niches

Back to glossary

Loading…