Skip to main content
Agents are the configurable AI workers in your workspace. In V3, the Agents area is where you create agents, decide what each agent can access, tune how it responds, and publish changes. Open Agents from the Agents group in the workspace sidebar.

​
Agent Types

​
Customer Agents

Customer agents are customer-facing chat agents. Creating one also creates an inbox, so inbound conversations for that contact type can be routed to the agent. Use customer agents for support, sales, guest messaging, lead intake and any other usecase where the AI responds to a contact.

​
Internal Agents

Internal agents run behind the scenes. They do not own an inbox and they do not respond directly to customers unless you connect them to another workflow. Use internal agents for operations work, data lookup, team-facing Slack workflows, enrichment, triage, and other tasks your team wants to automate out of the customer’s view.

​
Voice Agents

Voice agents are configured separately because they need phone-number routing, voice behavior, transfer rules, and call-specific guardrails.

​
Create An Agent

1

Open Agents

Go to Agents.
2

Choose an agent type

Click Add agent and choose Create customer agent, Create internal agent, or Build with Operator.
3

Name the agent

Use a short, unique name. Agent and inbox names must be unique in the workspace, and names are limited to 20 characters.
4

Configure the agent

Open the agent and fill the Persona, Tools, Skills, Behavior, and Settings tabs that apply to that agent type.
Use Build with Operator when you want Conduit to help create the agent from a plain-language description of the job it should do.

​
Chat Agent Tabs

Chat agents use these tabs: Internal agents do not show every customer-reply behavior setting because they are not inbox-bound response agents. For a customer agent, a run is logged in Activity each time a message arrives in the inbox that agent is listening to. Open the most recent run to confirm the agent is using the instructions, tools, skills, and sources you expect.

​
Persona And Guardrails

Persona and Guardrails both control how the agent behaves, rather than what it knows or what it can do.

​
Persona

Persona is the agent’s personality and general operating instructions. A new agent starts with defaults covering things like:
  • Which pronouns to use, for example we, our and us rather than I, me and my
  • How long a reply should be, for example a maximum of two or three sentences
  • When to check listing information first, for example any question about pricing or availability
Customise it to match how your team writes. Teams often add a local greeting, or specify Australian English rather than American English.
If you want the agent to phrase things differently every time, put it in Persona. Response style is not knowledge, so it does not belong in the Knowledge Base, and correcting the agent in a single conversation will not apply the rule everywhere.

​
Guardrails

Guardrails are the things the agent must always or never do. Common examples:
  • Never describe itself as an AI or a bot, and always present as a member of the host team
  • Never share one guest’s information with another
  • Never use offensive language
  • Never send a guest to a third party, such as telling them to contact the OTA directly
A guardrail against third parties stops the agent sending a guest elsewhere. It does not stop you building a skill that contacts a vendor on your behalf. If you want the agent to apologise to a guest about the pool and then message the pool cleaner, that is a skill instruction rather than a guardrail change.

​
Publish And History

Agent edits are staged before they affect production behavior. When the save bar appears, review the changes and publish them intentionally. Agents also keep version history so you can inspect previous published configurations. Use Activity after publishing to confirm that the agent is using the expected instructions, tools, skills, and sources.

​
Triggers

Customer agents can be attached to an inbox trigger. The agent sidebar shows the current inbox trigger and warns if another agent is already listening to that inbox. Internal agents can be connected to team channels. Slack is available in the internal-agent channel area. Other channel trigger rows may appear disabled until they are enabled for your workspace. Internal agents also support these trigger types:
When you are still writing a scheduled agent’s skills, create the schedule but leave it inactive. You can then build and test without the agent running against live data, and activate the schedule once the behaviour is right.
A Workflow trigger doesn’t require you to build a companion workflow. Pick the event the agent should react to — a new reservation, a label being added, a PMS reporting that a unit’s cleaning status changed, and so on — and the agent runs automatically whenever that event fires, anywhere in the workspace. Workflows and agents share the same underlying events, so the same trigger that starts a Workflow can start an agent instead.
Workflow or agent? If something should happen the same way every time, build a Workflow — treat it like code, deterministic and consistent. If the right action depends on judgment, or there are too many conditions to enumerate in advance, build an agent instead. A useful test: if a human would be inconsistent deciding what to do (say, right only half the time), that’s a sign the task needs an agent’s reasoning, not a fixed set of rules.
See Internal Agent Use Cases for worked examples, including a cleaning-status agent that reacts to a live PMS webhook today. Treat disabled trigger controls as unavailable until the product surface enables them in your workspace.

​
Behavior Settings

The Behavior tab controls how a customer-facing chat agent operates after it knows what to say.
Only edit Hooks when you understand the handler you are configuring. Hooks can call tools, HTTP endpoints, LLM handlers, or agents before and after an agent run or tool call.
Calling another agent is a Hooks-specific capability. The Skills and Tools composition described elsewhere on this page stays scoped to the agent that owns it — Hooks are the mechanism for one agent’s run to trigger another agent.

​
Legacy Agents

Older agents may show an Upgrade action. Upgrade behavior is migration-only context; new V3 setup should use Skills, Tools, and the current Knowledge Base. If your workspace still has legacy agents, complete the Knowledge Base migration before upgrading all chat agents.