Agents
The unit that combines persona, skills, knowledge, tools, entities, signals and policy into something you can govern.
An agent is a deployable conversational capability. It is not a prompt behind an API. It is a configuration that defines how the agent behaves, what it can reach, what it is allowed to do, and how it stays safe in an enterprise setting.
The agent is the design-time definition. A conversation is the runtime instance where that definition is actually used, by a person, in a meeting, a call or a chat, inside a specific context.
What an agent is made of
Persona
The agent's role, tone, objectives and conversational boundaries. What it is responsible for, how it speaks, what it should refuse or escalate.
Skills
Packaged behaviours that make outcomes repeatable: how it runs a discovery conversation, how it structures a summary, when it reaches for a tool.
Knowledge
The documents and approved context it grounds answers in, so guidance reflects policy rather than general reasoning.
Tools
Controlled access to other systems, so the agent can check rather than guess. Grouped into toolkits in the console.
Entities
What the agent works about: a customer, a quote, a case. Also the unit that access control and aggregation hang off.
Signals
What it should notice: decisions, intent, rationale, constraints, risks and commitments.
Actions
Controlled follow-through, so what was agreed in the conversation becomes a change in a system of record.
Policies
The constraints over all of the above: data scopes, tool restrictions, action boundaries, evidence requirements and privacy rules.
Persona
Persona is more than a system prompt. It sets role clarity, so the agent knows what it is and is not responsible for; interaction style, meaning tone and pacing; objective framing, meaning what it is trying to achieve; and boundary conditions, meaning what it avoids, escalates or refuses outright.
Skills
Skills are modular behaviours: how to assemble context before answering, how to write a structured recap, how to identify a signal and record it, how to draft a document to a house format, when and how to use a permitted tool.
Turning a skill on is a governance decision as much as a capability one. A skill should be traceable while it runs, permissioned so you know who can use it, and auditable afterwards.
Knowledge
Knowledge is how an agent stays grounded in your reality rather than in general reasoning: policies, playbooks, manuals, standard operating procedures. It reduces ambiguity, aligns answers to the official version, and keeps different teams getting the same guidance. Access is always scoped.
Tools
Tools expand what an agent can know. They retrieve authoritative data, confirm state, and assemble context under approved scopes. They are surfaced through toolkits so credentials, permissions and audit are handled centrally rather than being embedded in an agent's instructions.
Tools do not decide outcomes. They provide the access needed to be accurate.
Entities
People converse in the context of business objects: customers, quotes, cases, opportunities. Entities make that explicit. Conversations attach to them, access is commonly granted at their level, and intelligence aggregates at their level and rolls up from there.
Signals
Signals are how the organization's interpretation of a conversation is captured: what was decided, what trade-offs were accepted, what constraints shaped the outcome, what was committed to. They accumulate across conversations so continuity does not depend on anyone's notes.
Actions
Actions are the output side: creating and updating records, assigning owners, starting downstream processes. They are treated as first-class because execution has to be explicit, governed, safe to retry, and attributable afterwards.
Policies
Policies are the control plane. They set which data an agent can retrieve and from where, which tools it can invoke and under what conditions, what it may initiate, when an output must be grounded in evidence, and whether a given conversation may update shared organizational memory.
How people meet all of this
None of it is visible as a list of features. It is experienced as a conversation: someone joins a meeting, makes a call, or opens a chat, and the agent's persona, skills, knowledge, tools and policies come together against the specific context in front of them.
Where this lives in the product
Agents are created and configured in the console, on the Agents screen, and published there before anyone can pick them. On the phone, a published agent attached to a record's entity is what you choose when you start a call or a chat.