Niluxv1.0.33
/agents

Subagents.

For a big task, Nilux can hand parts of the work to subagents: helper agents that each work on one piece with their own conversation. For example, one studies how the login works while another writes tests. They work in parallel, so the whole task takes less time, and each one reports back when it is done.

You do not have to start agents yourself. Nilux decides when splitting the work helps. If you want it to, just say so: "use an agent to look through the billing module".

NOTE

Each agent uses its own tokens, so parallel agents finish sooner but cost more than one agent working alone. Use them for work that really splits into separate pieces.

Built-in agents

  • Explore: only reads and searches. Fast, and good for questions like "where is this handled?"
  • Task: a general helper that can also change files and run commands.

Agents follow the same approval rules as the main conversation; if one needs your approval, you see which agent is asking. Every agent except Explore also gets your project instructions.

Watch and steer agents

While agents run, a list appears under the prompt. Each row shows what the agent is doing right now, its model, and how long it has been running.

  • Press ↓ on an empty prompt, or Alt+A, to go into the list.
  • Press Enter on an agent to read its whole conversation. Press Esc to come back.
  • Type a message while an agent is open, and it is sent to that agent. A running agent reads it at its next step. A finished one picks up its work again with your message.
  • Press Ctrl+X to stop an agent. What it has done so far is kept.

A finished agent stays in the list for a minute, in case you want to tell it something more.

Agents in the background

Press Ctrl+B while agents are working to move them to the background. They keep working with the same tools, and you can keep talking to Nilux. When they finish, Nilux gets their results.

Your own agents

Create an agent for a job you repeat, for example a reviewer that checks code against your rules. The easiest way is /agents: choose to create one, describe what it should do, and Nilux can write the agent's instructions for you (Generate with AI). You then pick its tools, model and color.

An agent is a Markdown file in ~/.nilux/agents/ (all your projects) or .nilux/agents/ in a project (that project only):

markdown
---
name: reviewer
description: Reviews changed code against the team's rules
tools: Read, Grep, Glob
model: inherit
color: purple
---
You review code. Read the changed files, compare them with the rules
in nilux.md, and list problems with the file and line. Do not edit files.
  • name: the agent's name. It defaults to the file name.
  • description: when to use this agent. Nilux reads it to decide which agent fits a task, so be specific.
  • tools: what the agent may use: Read, Glob, Grep, Bash, Write, Edit, TodoWrite, WebFetch, WebSearch. Leave the line out to give it all of them.
  • model: inherit uses the same model as your conversation. In /agents you can set a different model per agent.
  • color: the color of the agent's rows: red, blue, green, yellow, purple, orange, pink or cyan.

The text after the second --- is the agent's instructions. Changes to the file apply immediately. If a project agent has the same name as one of yours, the project's agent is used.

Large jobs with many agents in planned stages are what Workflows are for.

← Previous · CLI referenceNext · Workflows →