Skip to main content

​
Overview

Agent Builder is a dedicated Copilot mode for designing, listing, inspecting, and configuring agents. Use it when you want to define an agent as a resource — instructions, applications, MCP servers, skills, and setup — rather than simply chat with one. Chat mode is for having a conversation (optionally with one agent selected via @). Agent Builder is for agent management: create, clone, validate, and wire integrations before you use the agent in everyday chat.
Agent Builder mode home screen

​
Chat vs Agent Builder

If you ask in Chat to create an agent, list your agents, or inspect full agent configuration, Copilot will suggest switching to Agent Builder and show a one-click Switch to Agent Builder card.

​
Switching to Agent Builder

​
From the home screen

On a new chat (home), use the mode toggle in the top bar:
  1. Chat — default conversational mode
  2. Agent Builder — agent creation and management mode
The message input border turns blue in Agent Builder so you can tell which mode you are in at a glance.
The Chat / Agent Builder toggle is available on the home (new chat) screen. From an ongoing thread, use the in-chat Switch to Agent Builder card or start a New chat to change modes.

​
From an ongoing chat

When Copilot detects agent-management intent (create, list, clone, inspect config, compare agents), it can show an Agent Builder shortcut card in the thread. Click Switch to Agent Builder to change mode without leaving the conversation context. You can also open Agents from the main menu (/agents, /agent/new, /agent/[id]) for the full agent editor UI.

​
What Agent Builder can do

Agent Builder uses specialized tools to help you end-to-end:
  • Discover applications — search APIs and MCP servers, browse a preloaded catalog of installed and official integrations
  • Generate and refine instructions — AI-drafted system prompts based on your use case
  • Skills — search your org library and skills.sh, propose skills to attach, and wait for your approval before creating the agent
  • Create and update agents — name, description, instructions, applications, MCP servers, contexts, and skills
  • Clone agents — start from an existing agent as a template
  • Validate configuration — check names, instructions, and attachments before saving
  • Present results — after creation, shows an agent card and setup instructions for connecting to Claude, Cursor, ChatGPT, and other MCP clients
Agent Builder follows a guided flow. You do not need to memorize tool names — describe what you want in natural language.

​
1. Describe the agent

Start with a short brief, for example:
Agent Builder may ask 2–3 clarifying questions about purpose, capabilities, and required integrations before proceeding.

​
2. Discover integrations

It searches your installed applications and official catalog, and recommends APIs (OpenAPI integrations) or native MCP servers as appropriate. Install native servers first via Native MCP Servers when needed.

​
3. Optional: attach skills

If your use case maps to reusable workflows (SEO, code review, deploy runbooks, etc.), Agent Builder can:
  1. Search your org Skills library and the public registry
  2. Show an interactive skill proposals card (1–5 matches)
  3. Wait for you to install, select, or skip — it never auto-installs without your choice
Skipped skills do not block agent creation. Agent Builder (and Chat) can also show a use case proposals card when you ask for suggested workflows across apps — review the proposals before acting on them. See Copilot → Use Case Proposals.

​
4. Generate instructions

Agent Builder drafts a structured system prompt (role, guidelines, boundaries) and can refine it based on your feedback.

​
5. Validate and create

Configuration is validated (name format, required instructions, valid app/MCP references), then the agent is created.

​
6. Review setup

Immediately after creation:
  • An agent details card appears in chat (click through to the edit page)
  • An agent setup card shows how to connect the agent to external AI clients via MCP

​
Agent naming guidelines

Agent Builder enforces concise names:
  • 2–3 words maximum (e.g. “Dev Assistant”, “Support Bot”)
  • Letters, numbers, spaces, hyphens (-), and underscores (_) only
  • Avoid long compound names like “Dev Flow Assistant (GitHub + Jira)“

​
Example prompts

Create from scratch
Clone and adapt
List and inspect
Skills-focused

​
After Agent Builder creates an agent

Fine-tune in the Agents editor:
  • Instructions — edit the system prompt directly
  • Applications — attach or remove OpenAPI integrations
  • MCP Servers — native MCP attachments
  • Skills — org skills library
  • Context — reusable knowledge items
  • Generative UI — visual templates for MCP App results (after the agent is saved)
  • Setup — MCP URLs, per-tool toggles, and managed setup export
See Agents for the full editor reference.

​
Best practices

Be specific about integrations
  • Name the systems you need (Slack, Jira, Gmail) so Builder picks the right apps
Answer clarifying questions
  • Short answers upfront save back-and-forth later
Use Chat for testing
  • After creation, start a new Chat thread and mention the agent with @ to try it in a real conversation
Iterate in the editor
  • Agent Builder gets you to a strong first version; use the Agents page for precise tool and instruction tweaks

​
Troubleshooting

​
I do not see the Agent Builder toggle

  • Open a new chat from the home screen — the toggle is on the new-chat view
  • From an existing thread, use the Switch to Agent Builder card or New chat

​
Agent Builder created an agent but tools fail in Chat

  • Open the agent in Agents → Applications / MCP Servers and confirm integrations are attached
  • Configure Security credentials for each application
  • Test with @agent-name in a fresh Chat thread

​
Skills were skipped — can I add them later?

Yes. Open the agent → Skills tab, or browse the Skills library and attach from there.

​
Next steps

  • Agents — edit instructions, tools, generative UI, and setup export
  • Skills — org skill library and registry
  • Copilot — use your agent in Chat mode
  • MCP Gateway — connect agents to external AI clients