Skip to main content
An agent is one LLM step in the workflow graph. 
agent summarize_task {
    model: model.fast
    instruction: "Summarize task {{ input.task_id }}."

    output {
        summary: string
        next_action: string
    }
}

​
Agent properties

PropertyPurpose
modelSelects a named model profile, such as model.fast.
usesMakes tools, prompts, and resources available to the agent.
contextContinues from another agent’s message history.
instructionThe instruction appended for this agent step.
fileUploads content before the agent call and injects it as a file attachment.
outputThe required structured output contract.

​
Uses

uses declares external capabilities and context available to the agent. It can include tools, prompts, and resources. 
agent writer {
    model: model.fast

    uses: [
        prompt.writer_instructions,
        resource.project_readme,
        tool.create_draft,
    ]

    instruction: "Create a draft for {{ input.topic }}."

    output {
        draft: string
    }
}

​
File directive

file uploads content before the agent executes and injects the resulting file attachment into the agent’s chat history. The content expression is positional and required; an optional block sets name and purpose. 
agent reviewer {
    model: model.fast
    instruction: "Review the uploaded file."

    file agent.summarizer {
        name: "report.json"
    }

    output {
        feedback: string
    }
}

​
File syntax

file <content> {
    name: "filename.ext"
    purpose: "file-extract"
}
The block { ... } is optional. When omitted, defaults apply.

​
File properties

PropertyRequiredDefaultPurpose
content (positional)Yes—An expression whose rendered value becomes the file body.
nameNo"file.txt"The filename exposed to the model.
purposeNo"file-extract"A label describing the file’s role in the workflow.
The content expression accepts references (agent.summarizer, dynamic.data), string literals ("manual content"), and number literals (132).

​
Content from an agent

file agent.summarizer {
    name: "report.json"
}

​
Inline content

file "manual content" {
    name: "notes.txt"
}

​
Multiple files

Multiple file directives are allowed on a single agent: 
agent compare {
    model: model.fast
    instruction: "Compare the two files."

    file agent.step_one {
        name: "before.json"
    }

    file agent.step_two {
        name: "after.json"
    }

    output {
        differences: string
    }
}

​
Agent loops

Use for ... in when each item in a collection should be processed with the same agent. 
agent summarize_each for task in input.tasks {
    model: model.fast
    instruction: "Summarize {{ task.title }}."

    output {
        title: string
        summary: string
    }
}
Loop headers may destructure object-shaped items: 
agent summarize_each for { a, b } in input.something {
    model: model.fast
    instruction: "Summarize A={{ a }} and B={{ b }}."

    output {
        summary: string
    }
}
The output of a looped agent is an array of that agent’s output objects.

​
Context continuation

Use context: agent.some_agent when a later agent should continue from an earlier agent’s message history. 
agent investigate_task {
    model: model.fast
    instruction: "Investigate task {{ input.task_id }} and identify the main issue."

    output {
        issue: string
        evidence: [string]
    }
}

agent propose_solution {
    model: model.fast
    context: agent.investigate_task
    instruction: "Continue from the investigation and propose a concrete solution."

    output {
        solution: string
        steps: [string]
    }
}
A structured reference such as agent.investigate_task.issue reads an output value. context: agent.investigate_task passes the prior conversation/message history so the next agent can append a new instruction and continue the work.