Skip to main content
Helix delegates parallel work to sub-agents: the model calls the Agent tool, the agent runs in the background, and the session keeps going. The messaging is bidirectional — the session can steer a running agent and read its result, and agents send findings back. You watch and steer all of it from two surfaces: the fleet list and the agent viewer. In the full Helix session (helix), Helix starts sub-agents itself when a stage needs parallel work. In agent mode, your agent has the same tools and you can ask for parallel work directly.

​
The fleet list

Dispatched agents appear under your prompt:
The fleet list under the composer: a finished agent with a checkmark, three agents running commands with live token counts, an overflow row reading + 1 more, and the key hint row.

The fleet list: a finished agent lingering, three running, one more counted.

Each row shows the agent type, a short label for its task, and live time and token counts. The list shows three or four rows, depending on your terminal’s height; the rest are counted as + N more, never dropped. The digits are slots held for an agent’s lifetime — an agent that finishes and ages out never renumbers the ones below it, so 5 opens what it opened a minute ago. Finished agents linger for a minute, and the clock restarts each time you open or close one, so an agent you keep returning to cannot vanish mid-read. The list steps aside while a picker such as /model holds focus.

​
The agent viewer

Open an agent and its conversation replaces the chat — it is not a split pane. The transcript renders with the same components as the main session: tool calls are real cards, markdown renders, diffs render, errors are styled as errors. It updates live while the agent works.
The sub-agent viewer replacing the chat: header with agent name and live stats, Bash and Read tool cards, the working line, and a steering instruction typed in the composer.

The agent viewer: live tool cards, the working line, and a steer typed into the composer.

The composer at the bottom is the same editor as the main chat, and whatever you type steers the agent you are watching. A steer that cannot land keeps your text in the box and tells you why — an agent that has already finished cannot be steered. Esc closes the viewer and returns you to the main conversation; the agent keeps running.

​
Resume a sub-agent

A sub-agent’s result ends with its id:
Helix can resume that sub-agent later in the session, with its earlier conversation intact, instead of starting a new one. Ask for it in plain words (“resume the explore agent and have it check the tests too”), or name the agent. Resuming, and sending a message to a sub-agent, accept any of:
  • the full id from the end of its result;
  • the short ref shown in the agent list;
  • a unique prefix of the id, at least four characters long.
A prefix that matches more than one agent is refused with the candidates listed, so you can pick one. The fleet list shows the new description when a sub-agent is resumed.

​
Which model a sub-agent runs on

A sub-agent runs on the model its definition names, or on the model of the session that started it. Helix can also ask for a size instead of a model: large, medium or small. When a size is asked for and none of its models is configured on your machine, or the model its definition names is not configured, the sub-agent still runs, on the session’s own model. You and the model both see a warning that names the model it runs on instead. A specific model asked for explicitly in the request, and not configured, is refused, never a silent swap.

​
Failures report as failures

An agent that never reached the model — no tokens, no tools, no output — is reported as an error with the reason (model availability, provider credential, rate limits), not as a finished agent with an empty transcript. A sub-agent that stops for inactivity reports that it failed, with how long it was inactive. A resumed sub-agent never returns the result of its previous run as if it were new.

​
Questions from background sessions

A Helix session can start another session in the background. When that background session asks a question, the question goes to the session that started it: The wait is 10 minutes when a person is there and 5 minutes when an agent answers.

The rest of the session surface

The dashboard, the session commands, and the working line.