Glossary

Every Cobalt term in one place.

Glossary

The core concepts an administrator works with, defined briefly. For how they fit together, see How Cobalt works.

Cobalt’s primitive nouns are capitalized as proper concepts — Agent, Skillset, Skill, Connection, Channel, Integration. Sensitivity values are lowercase: standard and elevated.

Agent

The thing employees talk to. An Agent has an identity (name, role, personality, safety boundaries, languages) and a set of tools it can use. Everything else you configure attaches to an Agent.

Skillset

A self-contained domain of custom capability that an Agent owns — a small chapter of related work. A Skillset groups Skills and references Connections, and carries a sensitivity. Use one only when an Integration can’t do the job.

Skill

A single playbook within a Skillset: plain-language instructions for one job, optionally including shell or curl. A Skill has a name, a “use when / skip when” description, a content body, and a requires confirmation flag for actions that change data.

Connection

The credentials a Skillset’s Skills need — named variables like a URL and an API key. Values are encrypted and injected at run time as environment variables (<CONNECTION_NAME>_<KEY>, upper-snake-cased). Secrets never appear in Skill text or in anything the agent says.

Integration

A Cobalt-built connector to an external system (service desk, HRIS, identity provider, file store, collaboration tool). An Integration turns that system’s capabilities into tools, with guardrails and audit built in. This is the primary way an Agent gains capability.

Tool

A single capability the Agent can invoke — a search, a lookup, an action. Tools come from Integrations, from Skillsets, or are built in. How a tool may behave is governed by rules and hooks.

Hook

A policy gate that runs at the moment a tool is called. A Hook can require confirmation, check an argument against an allowlist, rewrite a call, or deny it — with audit logging.

Custom knowledge

Reference material that doesn’t live in a connected system — an uploaded PDF or FAQ — organized into collections the Agent can search and cite. For content that does live in a connected system, prefer an Integration.

Channel

How employees reach an Agent: the web widget, Slack, and more. One Agent can serve many Channels. An Agent with no Channel is invisible.

Specialist

An internal helper scoped to a Skillset’s domain that handles multi-step work. Specialists are generated automatically from a Skillset’s Skills — there’s nothing to configure directly. To change one, edit the underlying Skills.

Identity Rule

A mapping from an inbound identity (an email, a Slack ID, a header) to a known person, so the Agent can act on their behalf and apply the right policy. Rules are evaluated in order; the first match wins.

Tenant

Your organization’s isolated workspace. Every object belongs to exactly one Tenant, and isolation is enforced at the data layer — one Tenant can never see another’s data.

Sensitivity

A Skillset is standard (default) or elevated. Setting elevated — for PHI, compensation, credentials, or security operations — routes work only to models covered by a Business Associate Agreement, tightens PII handling and retention, and isolates memory under a separate key.

Conversation

An ordered exchange between one or more users, the Agent, and (where used) a human during handoff. Conversations are what you review in Conversations & tracing.

Preview

An internal test surface for trying an Agent before connecting a real Channel. Preview conversations don’t reach employees and don’t count against usage.