Your first agent

An agent in Pumkin is a saved setup: a model, a system prompt that defines its behavior, and any tools you’ve given it. Once saved, you can run it over and over with different tasks. Let’s make one.

1. Create an agent

From the main screen, create a new agent. You’ll give it:

  • A name — anything. “Scratch”, “Research helper”, whatever you’ll recognize.
  • A model — pick from the models you’ve pulled into Ollama. If you followed the setup, llama3.2:3b is here. This dropdown is also how you swap models later to compare them.
  • A system prompt — the standing instructions that shape how the agent behaves on every run.

2. Write a system prompt

The system prompt is where you set the agent’s job and personality. Keep it concrete. A good starter:

You are a concise, careful assistant. When a task needs
current information or an action, use the available tools
rather than guessing. Explain what you did in plain language.

You don’t need tools configured yet to run an agent — without them it just reasons and answers. Tools are what let it act, and they’re covered next in MCP tools.

3. Run a task

With the agent saved, give it a task and run it. Try something simple first:

Summarize what you can and can't do right now.

Pumkin streams the agent’s work back as a run — you’ll see it think, respond, and (once you’ve added tools) call them. The whole exchange is saved so you can come back to it.

4. Read the run

A run shows each step the agent took, in order. Even on a plain question you’ll see the model’s response stream in. Once tools are involved, you’ll see tool calls and their results inline — this is the part that makes Pumkin an auditing tool, not just a chat box. You can see exactly what the agent did.

Here’s what a run looks like with a tool call (this is real output — llama3.2:3b using the built-in get_time tool):

$ ask "What time is it?"
[run_started]
[model_response] · tool_calls: get_time
[tool_call]   get_time({})
[tool_result] 2026-05-29T11:48:02.553Z
[model_response]
  "The current time is 11:48 AM UTC on May 29, 2026."
[done]

5. Keep going

That’s the core loop: configure an agent, give it a task, watch the run, refine the prompt. From here:

Next: MCP tools →