An AI node invokes an agent: a reusable, versioned AI job definition. Rather than configuring a model and prompt from scratch in every workflow, choose the agent that fits the job, provide its inputs, and state what it should do for this run.
Agents make repeated work consistent. Their instructions, output contract, model fallback, limits, tools, and permissions are defined once and can be reused across workflows. See Agents to create or customize one.
#Configure an AI node
- Add an AI node and choose an Agent.
- Fill in the inputs the agent requests. Required inputs are marked in the builder.
- Write the node’s Prompt: the specific task for this run.
- Optionally configure the standard retry policy.
For example, a reply-drafting agent might request a message, context, tone, and constraints. A workflow supplies those values, then the node prompt can say, “Draft a concise follow-up that asks for a meeting.”
The selected agent’s model, output type, tools, and estimated AI-credit cost appear as a read-only summary. To change them, open Manage agent. This keeps the recorded agent version an accurate description of what ran.
#Inputs, prompts, and variables
Use Insert data from an earlier step to pass information into an agent input or the node prompt. Choose the earlier step, then the detail you need—for example, an incoming email’s body, a lead’s company, or an earlier AI step’s category. Rills fills in the value before it invokes the agent.
See Use data from earlier steps for the picker walkthrough.
An agent’s own instructions use a different binding format, {{input.name}}. For example, an agent author might write Summarize {{input.content}} for {{input.audience}}. Keep those two layers separate: workflow builders provide values; agent authors define how those values are used.
#Use an agent’s output
An agent defines its own output contract:
- A plain-text agent offers one Text value in the picker.
- An agent with structured output offers each declared field separately, such as Category or Subject.
The variable picker shows the fields available from upstream AI nodes. Use structured output whenever a later node needs a specific value for a condition, approval, or action.
If you need a different output shape, duplicate the agent into your workspace and update its output contract. Do not parse text just to recover fields a workflow needs reliably.
#Versions and readiness
A workflow references an agent, not a hand-copied configuration. When an AI node executes, it uses that agent’s active version and records the exact version and model in the run details.
Before publishing an agent version, Rills checks changes that could affect workflows using it. For example, a new required input needs a default, and removing an output field can break a downstream binding. When a version becomes active, Rills rechecks the workflows that use it and shows any needed fixes in readiness.
Hidden or archived agents cannot be added to new workflow nodes. Existing workflow references continue to run, with a readiness warning so you can replace them deliberately.
#Retry policy and failures
AI nodes support the standard retry policy. If an agent reaches its output, time, turn, or credit limit before producing a complete answer, the node fails rather than passing incomplete output downstream. Increase the appropriate agent limit or adjust the task before retrying.
#Costs
AI calls consume AI Credits based on the model and tokens used. Agent run limits provide an upper bound before a call begins; the node summary shows an estimate for the selected agent.
Approvals, workflow logic, and time spent waiting for a reviewer are free. Web research may also use Workflow Credits when it calls a metered external service.
#See also
- Agents: create, customize, secure, and version agents
- Workflows: how a workflow fits together
- Runs & Versions: inspect what ran
- Billing & Credits: manage credit usage