Getting started

Quickstart: your first agent

Build a Sequential Agent with one input and one task, run it, and read its record. About ten minutes.

This walkthrough builds the smallest useful thing in Turtle AI Coworker: a Sequential Agent that takes one input, does one task and returns a recorded result. As an example we build aMeeting Notes Summarizer that turns raw notes into a short summary with action items. The same editor is used for every agent you will build later.

If you would rather not start from a blank page, install a Solution Pack or ask the Assistant to set something up for you.

Createname and instructions
Inputwhat the caller gives
Taskwhat to do with it
Runfill inputs, Run Agent
Create the agent, give it one typed input, write one task that uses that input, save, then run it from the Run drawer and open the record.

Step by step

  1. Open Agents and click Create New

    In the left navigation, under Build, click Agents. Check the workspace picker at the top of the page shows the workspace you want, then clickCreate New. (Import from Library is the other way in: it copies a ready-made agent template into the workspace.)

    app.turtleaicoworker.com/agents
    The Agents page with the workspace picker, Import from Library and Create New buttons, and agent cards
    The Agents page. Create New opens an empty agent in the editor.click to enlarge

  2. Fill in Identity and Behavior

    The editor opens on Basic Settings. Under Identity, type aName, for example Meeting Notes Summarizer, and a shortDescription. The description is shown on the agent card and in search, and it is how teams and AI employees know when to use this agent. Under Behavior, add anAgent Role (one line, for example Chief of staff who writes crisp summaries) and the Agent Instructions, the rules every task follows. For example: write in plain English, never invent a decision that is not in the notes, list each action item with an owner.

    app.turtleaicoworker.com/agents/new
    New Agent editor on Basic Settings with Identity and Behavior fields and an empty Workflow panel
    Basic Settings: Name and Instructions are required. The Workflow panel on the right is empty until you add a task.click to enlarge

  3. Choose the model

    Scroll down to Model Configuration. Pick the LLM Provider and theLLM Model from the ones your organization has connected. Set theTemperature: the slider goes from 0 (deterministic) to 2 (creative). For a summary, a low value such as 0.2 gives steadier results.

    app.turtleaicoworker.com/agents/new
    Model Configuration with LLM Provider, LLM Model and a Temperature slider
    Model Configuration: the provider, the model and the temperature for this agent.click to enlarge

  4. Add an input

    Open the Input Settings tab and click Add Input. InAdd New Input, set the Name to Meeting Notes, theKey to meetingNotes, the Type to Long Text, tickRequired field, and click Add. The key is how tasks refer to the value.

    Add New Input dialog with Name, Key, Type and a Required field checkbox
    Add New Input: a name people read, a key tasks use, a type and whether it is required.click to enlarge

  5. Write the task

    In the Workflow panel on the right, click Add task. Give the task a title. In Description, write the instruction and reference the input with its key in double braces, for example: Summarize {{meetingNotes}} in five bullet points, then list every action item with its owner and due date. In Expected Output, describe the shape of a good answer in plain text, for example: a heading Summary with five bullets, then a heading Action items with one line per item. Click Save task.

    A task being edited in the Workflow panel with Description and Expected Output fields
    A task: the instruction in Description, the shape of the answer in Expected Output.click to enlarge

  6. Save the agent

    Click Create Agent at the top right. The Unsaved changes badge next to it counts what has not been saved yet. Once the agent exists, the editor shows more tabs (Tables, Triggers, Statistics,Advanced) and a header with Run agent, History andSave changes. From now on, use Save changes after an edit.

  7. Run it

    Click Run agent. A drawer opens with the model, average runtime and cost, and one field per input. Paste some meeting notes into the input and click Run Agent. You can also run any agent from the Run page in the left navigation, which lists everything you can run in the workspace.

    The Run drawer showing the agent's model, average runtime and cost, and its input fields
    The Run drawer: fill in the inputs and click Run Agent. Required inputs are marked.click to enlarge

  8. Read the result and the record

    When the run finishes, open History on the agent, or Logs on its card, to see the run. Each run records the inputs, the output of every task, every tool call, the duration and the cost. On the Run page, Recent runs lists runs too; click one to open its trace. If the result is not what you wanted, change the instructions or the task, click Save changes and run again.

Field reference

The fields you used in this quickstart. The full list is on the Sequential Agents page.

FieldTypeWhat it does
Nametext, requiredHow the agent shows up across the workspace, search and runs. Up to 90 characters.
DescriptiontextShown on the agent card and in search results. Up to 280 characters.
Agent Roleone lineSets the agent's persona.
Agent Instructionstext, requiredThe rules every task follows. Markdown and {{ params }} are supported.
LLM Provider / LLM ModelselectThe provider and model this agent reasons with, from the ones your organization has connected.
Temperature0 to 20 is deterministic, 2 is creative. Keep it low for repeatable results.
Input: Name, Key, TypeinputA value the caller supplies at run time. Tasks reference it as {{ key }}. Tick Required field to refuse runs without it.
Task: DescriptiontextThe instruction for this step. Can use inputs and earlier task outputs.
Task: Expected OutputtextWhat this step should produce, described in plain text.

Related