​

Create your first agent

Creating an agent is a short guided flow in three phases: create (name it), profile (how it talks and what it won’t do), and capabilities (what it can reach). You start from a template, and nothing you choose is permanent — every choice is editable after the agent exists.

​

Choose a template

Templates pre-select the integrations a typical agent of that kind needs, so you start most of the way there: Pick the closest fit, then walk the steps below.

​

1. Name the agent

Give it a display name, a role, and a short description — the description is shown to employees, so write it for them (“Ask me about IT issues and requests”), not for the console.

​

2. Set personality

Give the agent a voice. Start from a preset and edit it freely:

• Direct & efficient — calm and competent; confirms before destructive actions; cites sources.
• Warm & explanatory — approachable; explains the why before the how.
• Terse — one sentence where possible.
• Formal — full sentences, professional register.

The preset just seeds a text box — make it yours. This maps to the agent’s Personality and related profile fields, which you can refine any time on the agent profile.

​

3. Set safety boundaries

Turn on the boundaries the agent must respect. Common ones include:

• Never repeat SSNs, credit cards, or unmasked PII.
• Don’t promise response times or SLAs.
• Hand off to a human if sentiment turns clearly negative.
• Refuse HR, legal, or compensation topics and hand off to a human.

Each boundary is either a hard rule (refuse outright) or a soft one (prefer to hand off). You can add your own.

​

4. Choose languages

Pick the languages the agent will hold conversations in, and which one is the primary fallback. You can revisit this on the agent’s profile later.

​

5. Choose integrations and authorize

The template proposes systems; keep, switch, or clear each, and add extras — a productivity suite, a collaboration tool, a file store, or your own MCP server. Review the access each integration requests. Some connect on behalf of each end user (the agent only ever sees what that user can already see); others need an administrator to grant access once. You can authorize now or finish the wizard and authorize after. See Integrations for how this works.

​

6. Add knowledge

Optionally upload reference material — policy PDFs, how-to docs, FAQs — that the agent can search and cite. This seeds the agent’s custom knowledge; you can add more any time.

​

7–8. Skillsets and live agent handoff

Two short explainers, nothing to configure here: what skillsets are (custom capabilities you can author later) and how handing a conversation to a human works. Both are set up from the agent’s Capabilities page after creation.

​

9. Review the toolset

A final review of every tool the agent will have on day one, based on your integration choices. Trim anything you don’t want live yet, then create.

​

Try it

Cobalt creates the agent and shows the status of each integration it connected. From here:

1. Open Preview and ask the agent a real question. Preview conversations don’t reach employees and don’t count against usage.
2. Refine its profile and capabilities until it behaves the way you want.
3. Connect a channel to put it in front of employees.

The agent improves as real conversations arrive — Cobalt surfaces where coverage is thin so you know what to add next.