Mindset docsBuildersBuildBuild an agent
Build an agent
Creates an agent with its prompt, model and resources, hands work between agents, and chooses where people can reach it.
Skills for your coding agent
Ask your coding agent#
Create the invoice exceptions agent for the finance team. Give it the supplier contracts knowledge base and the invoice comparison function, make its first version active, and keep it unavailable until I've tried it.What gets built#
An agent: a prompt, a model, and the resources it may use. Resources are operations on connections, functions, widgets and other agents. Each change is saved as a version, and a new agent starts unavailable, so only org admins can reach it.
| Prompt field | What goes in it |
|---|---|
| Purpose | What the agent is for, in a sentence or two |
| Personality | Its identity and behavior. The main free text of the prompt |
| Background | Domain and audience context |
| Policy rules | Rules it must follow. These are instructions to the model, and Mindset doesn't enforce them |
| Output formatting | How replies are shaped |
Each field holds at most 2,000 characters, and a longer one is refused, not trimmed. Mindset adds the reply language on every run, and the page's situational awareness when there is any.
What you decide#
- What stays in the prompt. Most of it. A rule that must hold every time belongs in a script condition or a function, or the operation stays off the agent.
- The model and reply language. Only models your org has a provider key for are offered. The reply language is Inherit, which follows the org, or one you name, such as "Spanish".
- The smallest set of resources. Anyone who can reach the agent can use everything it holds.
- Activate or stage. On an agent people already use, your coding agent stages each change as a version, and you activate it on Versions & Availability.
- Where it can be reached. The Hub, your own product, other agents, schedules, triggers or outside systems. Your coding agent never opens The Hub, Embedded in your own product or Outside systems on its own. How agents are started and reached explains each place.
Hand work to another agent#
A lead agent can hand a task to a helper agent and use its answer:
Add the supplier contracts agent as a helper on the invoice exceptions agent, and let other agents hand it tasks.The helper goes on the lead's Agents (A2A) resources, in a new version, and needs Other agents ticked on its own availability. The lead calls it as a tool named delegate_to__ and the helper's handle, for example delegate_to__supplier-contracts. The helper runs as itself, with its own resources and permissions, never the lead's. A chain can be several agents deep, but a hand-off to an agent already on the chain is refused, and the whole chain shares a budget of 8 hand-offs per run.
Done when#
- You talked to it on the Chat tab, or your coding agent called it, and it used the resources you gave it.
- One version is active, and you know which.
- It shows Available in the places you chose, and you can say in one sentence what it lets everyone there do. An Unavailable badge names what's missing.
In your repo#
The prompt fields are files in the agent's prompt/ folder, and its model, resources and availability are in agent.json. Your coding agent applies a change and stages it as a version. See Keep your agents in your code.
In the Studio#
Agents → + New agent opens the Agent Builder in a dialog. Describe the agent, then press Open agent →. Edit the prompt on System Prompt, the model and reply language on Settings, and what it uses on Resources. Press Save as version, try it on Chat, then activate it and choose where it's reached on Versions & Availability.