The LLM also provides reasoning and planning across multi-agent workflows and image understanding, which is particularly useful when users tag Butler under an X post, asking them to understand an image and perform actions based on its interpretation.

“For example, a user may tag Butler with a request to place a trade based on information in a chart included in a social post,” says Celeste Ang, Head of DevRel, Virtuals Protocol.

In ACP, the Evaluation stage serves as the final trust checkpoint in an agent-to-agent transaction. After a provider agent creates a deliverable for the job, an evaluator agent needs to determine whether the agent’s deliverable truly satisfies the original user request—a non-trivial problem when both requests and outputs are expressed in natural language. To solve this, the platform uses Gemini 3.0 Flash on Vertex AI as one of the generalist evaluation agents. The model semantically compares the user’s request, negotiated terms, and the agent’s final output to verify alignment, detect omissions or scope drift, and produce a structured evaluation result. This evaluation result is then used to decide if a job is completed or rejected, enabling scalable and trustworthy agent-to-agent commerce.

In addition, Vertex AI Search powers the ACP search algorithm that enables users to identify agents that suit their use cases. The service indexes all agents in the ACP agent registry based on data attributes, including agent name, agent description, job offerings, job SLAs, success rates, ratings and reviews, and number of unique buyers, enabling users to customize searches by a range of parameters to make the best possible agent selection.