Agents overview
Understand what an Infra One agent is, where it can work, and how its configuration, knowledge, team, and channels fit together.
An Infra One agent is an AI participant configured for an organization. It can talk with website visitors, work alongside your team, answer in connected Slack channels, and act as the controlled language layer inside an Automation.
The Agents service is the administrative surface for that identity. It owns the agent’s system prompt, model settings, teammates, knowledge sources, availability, avatar, and channel connections.
You need organization-admin access with permission to manage agents. If
agents.infra.onereturns a not-found page, the service may not be enabled for your organization or your account may not have access.
What an agent contains
| Part | What it controls |
|---|---|
| Name and avatar | The identity people see in agent-supported surfaces. |
| System prompt | The durable administrator instruction that defines role, boundaries, and behavior. |
| Model settings | The model, maximum output size, and response variability. |
| Team members | Organization members who automatically join every new website chat room for that agent. |
| Knowledge sources | Periodically fetched URLs added to the agent’s context. |
| Enabled state | Whether the agent may participate. Disabling is reversible. |
| Channel bindings | The website and Slack surfaces where the agent can be reached. |
A documentation-safe fixture of the current Agents page. Product labels and states match the live interface; account data is sanitized.
Where agents work
Website chat
Website chat uses a publishable embed key restricted to an explicit list of origins. The key identifies your organization and is safe to ship in page HTML only because the allowed-origin check constrains where it works. Team members assigned to the agent auto-join each new website chat room.
Slack
A connected Slack workspace can bind direct messages to one agent and individual channels to agents. A channel binding has a mode:
- Mentions lets the agent respond when it is mentioned.
- Off preserves the binding while stopping participation.
Automations can also use an enabled, channel-bound agent to post one generated message without waiting for a conversational mention.
Automations
The Agent message action is deliberately narrower than a chat. It runs one configured agent against one incoming JSON object, with administrator-authored trusted instructions, and posts the result to one bound Slack channel. The incoming object is evidence, not instruction.
Lifecycle and safety
Creating, disabling, and retiring are different operations:
- Create establishes the agent identity and initial system prompt.
- Disable pauses the agent without deleting its configuration.
- Retire removes the agent from every team and conversation it serves. Retirement requires typing the agent name and cannot be undone from the Agents page.
Disconnecting Slack is also different from deleting the Slack integration. A disconnect is reversible and keeps channel setup and mirrored history. Deleting removes the workspace, joined channels, and mirrored conversation.