A subject matter expert of internet safety, privacy, and security posted this about his¹ unexpected experience using Meta’s Muse agent to visit and browse the Amazon website². In the same thread, one may read this reply, and observe that the same agent might or might not achieve the same unexpected experience. After reading the post and the reply, one may wonder, “Why might an agent do that?”

To answer that question, let us define what an agent is. According to this post by a different subject matter expert of open source awesome LLM apps, an agent is a model³ and a harness⁴. As part of this setup, the model interacts with 3 types of harness components (1) a knowledge component⁵, (2) a controller component⁶ and (3) an execution component⁷.

Before an agent starts work⁸, a person must provide the agent a prompt, and then the agent starts doing stuff. During the process, an agent might do something unexpected such as this incident involving a company’s agent causing an outage in one of the company’s production environments. Therefore, to answer the question, “Why might an agent do that”, the answer may be classified into 1 of 3 categories:

To formulate a more actionable insight, one debugging an agent’s behavior may then dive deeper into each category’s subcomponents to decompose and then identify the root causes. To illustrate, when debugging a prompt, one option is to use this 6-part AI agent prompt framework. In addition, when debugging the model, another option is to review the configuration and/or parameter settings - and this assumes that most people are end-users of models. Lastly, when debugging a harness, an additional option is to study the 3 component layers and the corresponding subcomponent layers if a permission/allowance was granted⁹.

Returning to the case where a Meta Muse agent unexpectedly visited and browsed the Amazon website, firstly the person’s prompt is not included in the post and/or the post’s thread. As such, the prompt may have written a permission or an allowance for behavior that ideated and implemented ways to overcome a known constraint. Secondly, the agent’s model training process may have included information on ways to overcome specific types of constraints. Thirdly, the agent’s harness may have had at least one subcomponent which allowed and/or did not disallow for that behavior.

However, as one may observe, the probabilistic nature of agents makes it challenging to deterministically attribute unexpected behaviors to specific root causes¹⁰.