An agent configuration is a reusable contract. Define what the agent should do, what a workflow must provide, what downstream nodes can expect, and the boundaries for one invocation.
#Instructions and models
Write instructions that state the job, boundaries, and what a useful response includes. Reference declared inputs with {{input.name}}, such as {{input.accountName}}.
Choose models in fallback order and set creativity for the amount of variation the job needs. Rills tries the next model when the first is unavailable or blocked by workspace policy.
#Inputs and output
Inputs are values a workflow provides when it invokes the agent. Give each a lower-camel-case name, description, required state, and optional default. Defaults let you add a required input without breaking workflows already using the agent.
Define structured output when later nodes need reliable fields. A reply-drafting agent might return subject, body, and requiresHumanReview; a plain-text agent exposes one text result. Removing or narrowing an output field can affect every workflow that reads it.
#Limits, context, and permissions
Use Expert controls to bound one run:
- Research depth and exact limits cap turns, research calls, output tokens, AI Credits, and run duration.
- Context controls whether the agent receives no extra context, workspace business context, or workspace context plus remembered facts.
- Connected-app access can be No access, Read only, Selected operations, or Full access.
Grant only the tools and access the job requires. Broader access has a larger blast radius than editing instructions, so it requires administrator permission and explicit review.
#Publish safely
Saving a changed configuration publishes a new version. Before publishing, Rills identifies workflows affected by contract or permission changes so you can review the acknowledgement.
Improve an agent covers version history, feedback, and rollback.