Skip to content

Build your first agent

In this guide, you make an agent, test it, and save a fixed version.

What you need

  • GaugeDesk with a connected model provider
  • A small test project
  • A clear task that you can check

1. Create the agent

  1. Open the agent library.
  2. Select New agent.
  3. Give the agent a task-based name, such as README editor.
  4. Open an edit chat.

2. Write the instructions

Tell the agent:

  • what work to do;
  • what input it needs;
  • what output to produce;
  • when it must stop or refuse;
  • which tools it can use; and
  • how a person will check the result.

Use short instructions that you can test. Avoid broad instructions such as “Use any available data.”

Example:

Correct spelling and grammar in Markdown files. Keep the meaning, headings, links, and code examples. Do not edit source code. List each changed file.

3. Add only the tools it needs

For each tool, set:

  • the actions it can perform;
  • the files or services it can use;
  • input and output limits;
  • network access; and
  • what happens after an error.

Adding a tool does not give a run permission to use everything. GaugeDesk still checks the permission for each run.

4. Select the model provider

Select the provider and credential type. Do not put a credential value in the agent.

The provider receives prompts and allowed files in plaintext. Confirm that the provider is approved for the test data and future work.

5. Set the limits

Choose:

  • where the agent can run;
  • whether it can use the network;
  • which types of files it can use;
  • what output it can create;
  • how long hosted data can remain; and
  • what a published website agent may collect, if anything.

Start with narrow limits. Add access only when a test shows that the agent needs it.

6. Test the agent

Add the draft to your test project. Test:

  1. a normal task;
  2. a task with missing input;
  3. a file that the agent may not use;
  4. a tool that the agent may not use;
  5. cancellation;
  6. a model or network error; and
  7. an attempt to change its own instructions.

Review the transcript, file access, tool calls, denied actions, and changed files after each run.

7. Fix the draft

Use the edit chat to correct the instructions or limits. Repeat the tests until the result is predictable and easy to review.

8. Save a fixed version

Save a fixed version after the tests pass. Packages and website releases use a fixed version, not a draft.

You are done when you can identify the saved version and repeat the test with the expected result.

Next