C3 AI Documentation Home

Overview

Sooner or later your team repeats a task with the built-in C3 Agent, and repeats the explanation that goes with it: the same context typed into every conversation, the same procedure recalled from memory, the same follow-up questions to steer the agent back on course. Building an agent moves that knowledge out of the conversation and into a configuration, so it applies automatically every time.

An agent is a named, reusable configuration that shapes how a conversation behaves: the instructions it follows, the skills it can call on, the model it runs on, and the C3 applications it may connect to. Define one once, and everyone on your team chats with it as-is. Say your reliability team keeps fielding the same questions about asset alerts. Build a Reliability RCA Agent that already knows how to investigate one, and every engineer chats with that agent instead of re-explaining the investigation each time.

This page covers the parts of an agent, what it can connect to beyond its own workspace, and the order to build one in.

The parts of an agent

An agent bundles six parts: name and description, instructions, skills, model, default apps, and icon. You set them when you create and configure the agent; everyone else sees them on its detail page. See Create and Configure an Agent.

  • Name and description: the agent's identity in the gallery. A unique name, and a one-line summary that tells a colleague when to reach for it.
  • Instructions: what you tell the agent about its role, its priorities, and the rules it must not break, delivered at the start of every conversation as its AGENTS.md. See Write Agent Instructions.
  • Skills: instructions for one repeatable task, attached to the agent so it can use them when relevant. Create a skill when you find yourself giving the agent the same multi-step instructions over and over. See Create and Attach a Skill.
  • Model: the default LLM the agent's conversations run on, chosen from the models your administrator has enabled. A user can override the model per turn in chat.
  • Default apps: the C3 applications the agent is tied to. Leave this empty and the agent is unrestricted: whoever chats with it picks an application at chat time.
  • Icon: a picture you choose from a built-in set of icons, so you can spot the agent's card at a glance in the gallery.

Two of these carry most of the weight. Instructions govern how the agent behaves in general. Skills govern what it knows how to do in particular. A useful agent is usually a short, clear brief plus a small set of well-chosen skills, not a long brief trying to do the work of both.

Connections: what an agent can reach beyond its own workspace

An agent runs in a workspace of its own. A credential, an application on another cluster, and a third-party service all sit outside that workspace, so reaching any of them takes an explicit connection. Three kinds exist, and most agents need none or one.

  • Connect Secrets: add one when a skill or script needs a credential, such as an API token, a webhook URL, or a connection string, without it ever appearing in a prompt.
  • Connect Apps on Other Clusters: add one when the agent must query a C3 application that lives on a different cluster.
  • Connect MCP Servers: add one when you want the agent's tool catalog to extend to an outside service such as Jira or GitHub, with each user signing in as themselves.

Clone an agent instead of starting from scratch

Every environment ships with the built-in C3 Agent: a general-purpose assistant for data analysis, application inspection, and operations against live C3 applications. It arrives preloaded with the core skills and no default apps, so it works against any application you can reach the moment you connect one.

C3 Agent is also protected: read-only, so nobody can edit or delete it, which keeps a known-good baseline available to everyone. To build on it, clone it instead of trying to change it. Cloning copies the source agent's instructions, skills, model, and default apps, with you as the new author. Your first custom agent then starts from a baseline that already works, and you only have to describe the difference. Start from scratch instead when your agent has little in common with anything that exists. See Start from scratch or clone.

The order to build an agent in

Build these five steps in order, from a bare agent to one ready to hand to your team. Steps 1 and 2 are required; take the rest as your agent's job demands.

  1. Create and Configure an Agent: create the agent from the gallery, either cloning an existing one or starting from scratch, then set its model and default apps. Creating and configuring are one continuous flow — the Create dialog asks only for a name and description before dropping you into edit mode.
  2. Write Agent Instructions: give the agent its instructions. This is the highest-leverage part of the build, and the part most worth revising after you watch the agent work.
  3. Create and Attach a Skill: write a SKILL.md for each repeatable task, package it, upload it, and attach it. Write a skill when a procedure has real steps worth getting right. Leave it in the instructions when a sentence covers it.
  4. Share a Skill Across Your Team: promote a skill from Personal to Global so the whole cluster can attach it. Administrators only, and worth doing once the skill has proved itself.
  5. Connect what the agent needs: Connect Secrets, Connect Apps on Other Clusters, or Connect MCP Servers, as required.

Then test it in a conversation, which is the only way to find out whether the instructions and the skills you wrote land. See Have your first conversation. Expect to revise the instructions once or twice after watching real transcripts — that revision cycle is a normal part of building an agent.

After the agent behaves, run it without a person present with Create and Schedule an Automation.

Before you start

Three role checks decide what you can build and connect:

  • You need the Agentix.AgentAdmin role to create, edit, or delete an agent. Without it, the Create Agent button stays disabled and Edit never appears on a detail page. You can browse agents and chat with them, but not author them.
  • Any user can write and attach personal skills. Uploading a skill at Personal scope and attaching it to an agent which you author, only needs Agentix.User. Promoting a skill to Global requires an administrator.
  • Some connections are administrator-only. Personal secrets are open to every user to store and manage. Registering an MCP connection, personal or global, and setting up a cross-cluster token both require Agentix.AgentAdmin.

For the full role ladder and how an administrator grants a role, see Roles and Permissions.

Where to go next

These pages cover each build step in detail.

Was this page helpful?