06 / Agent creation

The agent definition stays readable.

The JSON defines what the model should return, then hands the repeatable work to a Cargo-AI tool.

The tool is a custom Rust binary behind the Cargo-AI contract. It can be tiny, or it can become a super tool with as much deterministic capability as the workflow needs.

Input

One photo

The agent handles a single animal photo per run so the workflow is easy to inspect.

Schema

Six fields

image_name, animal, count, confidence, garden_risk, and notes.

Action

record_finding

When the model has a finding, the action calls the Cargo-AI tool.

Tool

CSV super tool

The tool creates the CSV, checks headers, and appends durable rows.

animal-patrol.jsonAgent definition

{
  "version": "2026-05-19.r1",
  "inputs": [
    {
      "type": "text",
      "text": "Inspect exactly one runtime-supplied animal photo and return one Animal Patrol finding. If the caller supplies a runtime text hint in the form \"Image name: <file>\", copy <file> exactly into image_name. Otherwise set image_name to the base file name of the supplied image when the image path is available. Identify the animal most relevant to a garden patrol; use animal \"none\", count 0, confidence 0, garden_risk \"none\", and brief notes if no animal is visible. Use garden_risk only from: none, low, medium, high, unknown. Keep notes concise and presentation-friendly."
    }
  ],
  "runtime_vars": {
    "output_csv_path": {
      "type": "string",
      "default": "outputs/animal-patrol.csv"
    }
  },
  "agent_schema": {
    "type": "object",
    "properties": {
      "image_name": {
        "type": "string",
        "description": "Base file name of the inspected image."
      },
      "animal": {
        "type": "string",
        "description": "Animal identified in the image, or none if no animal is visible."
      },
      "count": {
        "type": "integer",
        "description": "Number of animals identified in the image.",
        "minimum": 0
      },
      "confidence": {
        "type": "number",
        "description": "Confidence from 0.0 to 1.0.",
        "minimum": 0,
        "maximum": 1
      },
      "garden_risk": {
        "type": "string",
        "description": "Estimated garden risk.",
        "enum": ["none", "low", "medium", "high", "unknown"]
      },
      "notes": {
        "type": "string",
        "description": "Short notes supporting the animal and garden risk assessment."
      }
    }
  },
  "actions": [
    {
      "name": "record_finding",
      "logic": {
        "!=": [
          { "var": "image_name" },
          ""
        ]
      },
      "run": [
        {
          "kind": "tool",
          "name": "animal_finding_recorder",
          "params": {
            "image_name": { "var": "image_name" },
            "animal": { "var": "animal" },
            "count": { "var": "count" },
            "confidence": { "var": "confidence" },
            "garden_risk": { "var": "garden_risk" },
            "notes": { "var": "notes" },
            "output_csv_path": { "var": "runtime.output_csv_path" }
          },
          "output_variable": "record_result"
        }
      ]
    }
  ]
}

Cargo-AI tool

animal_finding_recorder

A project-local Rust binary with a typed interface and owned CSV behavior.

Creates fileValidates headersAppends rows