Agents and workflows.

Nilux can hand parts of a larger task to helper agents. Each agent has its own conversation, and several can work at once. In the app you see every agent at work, and you can read it, write to it, stop it or move it to the background.

In the conversation

Each agent appears as a card in the conversation: its kind, what it was asked, its model, how many tools it has used, what it is doing now and its status. Click a card to open the agent.

While Nilux waits on its agents, the line above the message box counts how many are working. Move to background (Control+B) lets them go on in the background so you can write to Nilux again.

The Agents panel

The Agents chip in the session's bar, or Control+G, opens the panel: Main conversation first, then each agent with its status, its model and the step it is taking. A finished agent stays in the list for about a minute.

• Click an agent to open it. Main conversation or Esc brings you back. Opening an agent never stops it.
• Control+X stops the selected agent, as its Stop button does.
• An agent waiting for your approval is tinted in the list, and the approval card names it.

An agent's view

An opened agent takes the conversation's place: the task Nilux gave it and every step it takes. The message box then writes to that agent. A message to a running agent shows Queued, then Delivered; a message to an agent that stopped or failed takes it up again from where it was. When it finishes, its answer goes back to the main conversation. Back to main returns to the conversation, Stop stops the agent.

Background work

A command, an MCP call or the agents can go on in the background while the conversation goes on: press Move to background on its line, or Control+B. A count of what is running appears above the message box (for example 1 running task); click it for Background tasks: what is running and what has finished, with Stop this task. If work is still running when you quit, Nilux asks first.

Your own agents

Settings → Agents (or /agents) lists your own agents: personal ones, which work in every project, and those of the project selected in the sidebar. They are files: personal agents in ~/.nilux/agents, a project's in its .nilux/agents. The terminal's /agents edits the same files.

New agent creates one. Describe what it should do and press Write it: Nilux fills in its name, description and instructions for you to read (this needs a running session). Or fill them in yourself:

• Where it lives: only in the selected project (shown by its name), or in every project (Personal).
• Name: lower case letters, digits and dashes. It is also the file's name.
• Model and Colour.
• Description: when to hand it work. The session reads this to decide.
• Instructions: what it is told before every task: its role, how it works, what it reports back.

Then press Create agent. Open an agent to change it later. There you also choose its Tools (Every tool, or only those you tick) and whether it may run in the background (Background runs; off, a session always waits for it to finish). Save or Command+S saves.

Under Built-in agents, choose the model of Explore agents (read-only scouts) and General agents (agents that can change files), or keep Same as the session.

Workflows

A workflow is a team of agents working on one large task in phases: for example, review every module, then fix what the review found, then check the fixes. In a session, start your message with /workflow and describe the job:

Nothing starts until you agree. Nilux plans the run and shows Start workflow?: how many agents it will use (and your limit), the phases, the models, and what the agents may do: only read, or write in the listed paths and run the listed commands. Edits outside those paths still ask you, and deleting files is never allowed. A workflow runs many agents, so it can cost many times an ordinary reply. Start (Enter) runs it, Cancel (Esc) does not.

The run works in the background. Its progress is above the message box: how many agents are done and which phase it is in. Click it, or send /workflow alone, to open its board:

• each phase, and each agent with its status; click an agent to read its transcript;
• Stop run ends the run; Back to chat (or Esc) returns to the conversation.

An agent that fails is tried again, up to three times. If it still fails, no new agents start, those at work finish, and the run pauses: Resume goes on, Finish run ends it as partial. If the app closes during a run, its board offers Run again: finished agents are kept, not run or paid for again.

When the run ends, Nilux writes its report to a file instead of pasting it into the conversation; Open report shows it in the Report tab beside the session. Past runs on a board lists the session's earlier runs.

Settings → Workflows sets the Workflow agent limit (the most agents one run may start, up to 50) and the Workflow model (what the agents use when the request names none). Workflow settings on the start card changes the same two.