Create an agent
The builder walks through eight steps. Steps are listed on the left with a summary of what you set, the Overview panel on the right shows the agent at a glance and the checks left before you can publish, and the Assistant tab drafts or changes the agent for you.
Nothing is saved until you choose to. Save draft keeps the agent private to you; Save & publish agent runs the checks and publishes a version. More has Read the definition and Check definition. If you leave with unsaved changes, the builder asks before it discards them.
1. Purpose
- Agent name and Desired outcome: the outcome is one sentence describing a useful result. It is shown beside the name.
- System prompt: how the agent works, its role, the sources it uses and when it asks for help. Write with the assistant drafts it for you. Tool access and approvals are set in later steps, not here.
- First task: a realistic request to try the draft with before you publish.
2. Models & team
- Lead model writes every answer. Switch between Your models (providers your organization connected to the LLM Gateway) and Quilr models. Add provider opens provider setup if the model you want is missing.
- Team models: up to four extra models that specialists can use. The lead still writes the final answer.
- Specialists work in order and share their findings; the lead combines them into one answer. Add one when a second opinion helps, such as a reviewer that checks the lead's findings.
The engine options in step 7 adjust so that the engine works with every model on the team.
3. Integrations
Pick a Connection (any MCP server on your MCP Gateway), then choose which of its tools the agent may call. Only the tools you tick are available to the agent.
- Read-only tools are selected when you add a connection. Tools that change data are your call.
- Filter with All, Read-only and Changes data, search tool names and descriptions, or use Select all, Read-only only and Clear.
- If you have not connected your own account to the server yet, select Authorize (for example Authorize Atlassian), then Refresh and select the tools. Until then the Your tools are connected check does not pass.
- Add Slack opens the MCP Library on the Slack server so you can add Slack as a tool. See Slack as a tool.
- Browse integrations marketplace and Manage connections open the MCP Gateway library and your connections.
Without a tool, an agent can only answer from the model and never act.
4. Capabilities
Extra reach beyond your connected tools. Each one runs through your MCP Gateway and is listed for the people you share the agent with.
5. Skills
A skill is a written playbook, such as how to review evidence or format a report. Attach up to eight from the Skills Library: search, review a skill's instructions, then attach it. Manage skills opens the library.
A skill's instructions are copied into the agent version, so later library changes never alter a published agent. Skills are optional.
6. Inputs & output
- What people fill in: every run starts with a task in the person's own words. Add input adds a field for anything the agent needs every time, such as a customer name or a date range. Each input has a label (what people see), a name (how the agent refers to it) and a Require switch. What people will see previews the start form.
- What it always knows: variables you set once, such as a GitHub organization or a default repository. Write
{{name}}in the instructions and the value is filled in when the agent runs. People running the agent cannot edit them. They are part of the instructions, not secret storage, so do not put credentials in them. - What it hands back: the Output format, enforced on every run.
7. Runtime
The engine carries out the agent's steps and tool calls. One that works with every model you picked is chosen for you; change it only if your team prefers a specific framework.
8. Access & limits
Run limits stop a run when it reaches either one:
- Maximum turns: one turn is one model reply, including the tools it calls. 20 suits most tasks.
- Run timeout (seconds): defaults to 300. Time spent waiting for tool approvals counts toward it.
Who can run the agent, and whether their runs wait for approval before tools act, is decided when you share it. See Share and manage. Model and tool access are checked again on every run.
Publish
Before you publish lists what still stands between the draft and a published agent, for example It has a name, It has instructions, A model is chosen and Your tools are connected. Worth doing, not required suggests improvements such as adding an example task. When every check passes, Save & publish agent publishes a version.
Related
- Library templates: start from a pre-filled agent
- Build a team: combine published agents
- Share and manage