Start from the library, not from a blank prompt
Most agent projects die in the design phase because the first question is too open. Ivorycom's agent library gives the common revenue roles as working starting points, each with a defined objective and a scoped set of tools, so the first decision is which job to delegate rather than how to build an agent from nothing.
- ✦Lead Qualification — works new inbound against your fit criteria and hands over what is worth a person's time.
- ✦Account Research — assembles context on a company and the people in the deal before the call.
- ✦Deal Coach — reads the opportunity and surfaces what is missing or stalling.
- ✦Renewal — watches the signals that precede a renewal conversation going badly.
- ✦CPQ — assembles pricing and quotes from the deal context.
The agent library and Build-an-Agent
What you configure when you build one
A custom agent is defined by five things, and each of them is a control surface rather than a cosmetic setting. Identity determines how it presents itself. Instructions define the objective and the constraints. The model tier is chosen per workspace from your provider's live model list rather than pinned by us. Skills determine which tools it may call. Channels determine where it is allowed to speak.
- ✦Identity — the name and role the agent operates under, so a log entry is attributable.
- ✦Instructions — the objective, the tone, the escalation rules and the things it must not attempt.
- ✦Model — selected per workspace, including your own provider keys for Anthropic, OpenAI, Google or Grok.
- ✦Skills — the specific tools the agent may invoke; anything not granted is unavailable, not merely discouraged.
- ✦Channels — where output may go, which is what keeps a research agent from emailing a customer.
Least privilege, enforced rather than requested
There is a meaningful difference between an instruction telling a model not to do something and a system that will not let it. Prompt-level constraints are advisory. In Ivorycom, an agent's tool access is granted explicitly and every invocation is authorized against the workspace's policy before it runs — so an agent that has not been given a capability cannot reach it by being asked cleverly.
- ✦An agent receives the narrowest capability that does its job.
- ✦Every tool call is checked against your tenant's configuration and hard limits before execution.
- ✦If the policy service cannot answer, the action is denied — it does not fall through to permitted.
- ✦Every step, tool call and external request is written to an append-only record.
Roles, audit and policy
Supervision is a surface, not a promise
Agents that run continuously need somewhere to be watched. The command center streams runs, decisions, approvals, errors and workflow health into one view, so an operator can see what ran, what it decided, what it was refused, and what is waiting on a human. Sensitive actions pause there with full context rather than executing and notifying.
- ✦Live view of agent runs, automations and system events.
- ✦Approval queue where consequential actions wait with the reasoning attached.
- ✦Failure inspection, so a bad run becomes a tuning input instead of a mystery.
The AI command center
Giving agents access to the rest of your stack
An agent confined to the CRM can only do CRM work. Ivorycom uses the Model Context Protocol so agents can reach external tools through a governed connection, alongside native integrations and the Zapier catalog. The access is granted per workspace and passes the same policy check as any internal action.
- ✦MCP servers extend agent reach to external systems under the same guardrails.
- ✦Native connectors cover email, communication, ads and finance tooling.
- ✦Bring your own provider keys so model usage runs on your account and your terms.
How integrations and MCP fit together
Frequently asked questions
What is an AI agent in a business context?
A software worker with a defined role, an objective, a bounded set of tools and a supervision trail. It differs from an assistant in that it starts work on its own when a condition it watches changes, and from an automation in that it chooses the action rather than replaying a branch someone drew.
Can we build our own agents?
Yes. Build-an-Agent sets the identity, instructions, model, skills and channels. Skills are the access control: an agent can only call tools it has been granted.
How do you stop an agent doing something harmful?
Tool access is granted narrowly, every invocation is authorized against your workspace policy before it runs, a failed or unavailable policy check denies the action rather than allowing it, and high-risk categories such as money movement and bulk deletion require explicit human confirmation.
Which model do the agents use?
The model tier is chosen per workspace from the provider's live model list, and workspaces can run on their own Anthropic, OpenAI, Google or Grok keys. Model identifiers are not hard-coded by us.
Can agents use tools outside Ivorycom?
Yes, through Model Context Protocol servers and native integrations. That external access is granted per workspace and passes the same policy check as internal actions.
How do we see what an agent actually did?
Runs, decisions, approvals and handoffs appear in the command center, and the underlying audit record is append-only, so the history of a run cannot be edited after the fact.
Ivorycom