How do AI agents differ by their interaction mode?

Salesforce's Agentforce architecture whitepaper, aimed at technical architects building enterprise AI agents on its platform, sorts AI agents into five types defined by "their primary mode of operation and interaction": conversational, proactive, ambient, autonomous, and collaborative. Conversational agents "operate in a reactive, request–response manner," serving as "the digital front doors to an organization." Proactive agents, by contrast, "are triggered by specific events, data changes, or predefined conditions" and "execute a task or initiate a workflow without requiring direct user interaction." Ambient agents are "a specific type of proactive agents that operate continuously in the background of a user's workflow without requiring explicit commands," where users "benefit from their actions without being consciously aware of their operation." Autonomous agents are "given a high-level goal" and can "independently plan[] and execut[e] a sequence of steps," even "learn from their actions to improve performance over time." Collaborative agents, or "agent swarms," are "a collection of specialized agents that work together," coordinated by an "orchestrator" that "decomposes a large task, delegates subtasks to the appropriate specialized agents, and then synthesizes their outputs."

The whitepaper frames the five types as an ascending scale of initiative and complexity rather than a flat list: conversational agents wait to be addressed; proactive and ambient agents act on triggers without being asked; autonomous agents are delegated an entire goal and must plan the steps themselves; collaborative agents distribute that planning across multiple specialized agents coordinated through "a robust Agent2Agent (A2A) communication protocol." Salesforce justifies ambient agents as a way to "reduce the cognitive load on users by automating the mundane 'work about work'" by capturing and structuring information without an explicit request, and justifies collaborative swarms as mirroring "a human team" in which each agent can "bring its unique skills to bear."

The taxonomy runs parallel to, but is not identical with, Does machine agency exist on a spectrum rather than binary?: Rammert's spectrum sorts machines by degree of agency (passive to cooperative) independent of product category, while Salesforce's five types sort by trigger and interaction channel inside a CRM product line, with "collaborative" naming a multi-agent orchestration pattern rather than Rammert's highest-agency single machine. The whitepaper's proactive, ambient, and autonomous categories also complicate Why can't conversational AI agents take the initiative?: where that note finds initiative-taking systematically absent from LLM-based conversational agents, Salesforce's architecture defines event-triggered and goal-delegated agent types specifically to supply that missing initiative — though this is vendor documentation describing intended design, not an evaluation of deployed behavior. And where Do LLM agents misinterpret delegated requests more than formal protocols? argues that agent-to-agent meaning should be fixed by formal protocol rather than inferred statistically, Salesforce names its own "Agent2Agent (A2A)" protocol as the coordination mechanism for collaborative swarms without specifying whether it carries formal commitment semantics or is itself a thinner, LLM-mediated handoff.

The excerpt is architectural and promotional rather than empirical: it defines categories and asserts their business value — fewer escalations, lower service costs, a "true digital employee" — without reporting measured outcomes, adoption figures, or comparisons across deployed agents. It does not establish that the five categories are mutually exclusive in practice, or that Agentforce implementations of "autonomous" or "collaborative" agents actually exhibit the planning and learning capacities claimed; the taxonomy is best read as Salesforce's own product-architecture vocabulary for structuring its agent platform, not as a validated model of agent behavior.