Agents

Your agent doesn’t have to do everything itself. It can hand part of a job to a subagent: another agent with its own instructions and tools, which takes one task, does it, and reports back. This screen lists every subagent this workspace can delegate to.

Where to find it

Abilities → Agents in the sidebar.

What’s on the screen

The context line

Under the Agents heading, one line tells you what your roster costs, for example “4 on · about 180 tokens of context every turn. Switch off the ones you do not use.”

Here’s why. So the agent knows who it can delegate to, a short line about each subagent that’s on goes into its instructions on every turn. That takes up a few tokens (the unit models read and bill text in, roughly four characters) of its context, the text the model can see. A subagent you turn off costs nothing.

The list

Each row shows:

• on or off. A row that’s off is dimmed.
• The agent’s name.
• Where it comes from, in this order on the list:
  • project: this project’s own agents, kept in its .pi/agents (or .agents/agents) folder.
  • bundled: agents that ship with HappyVibe.
• The model it uses, if it has one of its own.
• What its line costs, for example ~40 tok: “What this agent’s line costs in the system prompt, every turn”.
• What it’s for, and the tools it may use.

Click a row to open it. You see its Tools it may use, its full System prompt (the instructions it works from), and the buttons Duplicate, Edit, Enable or Disable, and Close.

Before your first session

The list comes from a running agent, so on a fresh install it’s empty: “No agents yet” and “The bundled ones appear after your first session runs.” Start a session and come back.

Turn an agent off

1. Click the agent’s row.
2. Click Disable. Hover it first to see how much you save: “frees about” so many “tokens every turn”.

It takes effect from your next message. No restart needed. The agent stays on the list, marked off, and Enable brings it back whenever you like.

Edit an agent

Edit appears for bundled and project agents, the ones HappyVibe can write to.

1. Click the agent’s row, then Edit.
2. Pick a Model (agent tier), or leave it on Inherit (global / workspace default) to use your default model.
3. Change the System prompt if you want the agent to work differently.
4. Click Save. The change “applies to the next delegation”.

Duplicate an agent

To try a variation without touching the original:

1. Click the agent’s row.
2. Click Duplicate.

A copy named name-copy appears on the list, next to the original, in the same folder. Edit the copy, and turn the original off if you only want the new one. If the copy can’t be made, the window stays open and says why.

During a session

In a session’s top bar, a chip counts the subagents you can delegate to, for example 4 agents. Agents you’ve turned off aren’t counted. Click it to see each one with what it’s for. Picking one writes “Ask name to” into your message, so you only have to finish the sentence.

You can also type @ in the message box. Matching agents appear above your files, each as @name. Click one, then write the task, and the agent takes it as a request to hand that task over.

While a subagent is running, @ also offers it. A message sent that way “sends your message to this running sub-agent, not to the main agent”.

Related

• Agent tools: everything the agent can call, and whether it’s allowed to.
• Models: the default model an agent inherits.
• Skills: extra know-how the agent loads for a task, without handing it off.
• Prompts: reusable messages you send with /name.