Skip to main content

Runs

A run is one conversation with an agent: your messages, the agent's responses, and every tool call in between. If you've used Claude Code or Codex, it's what those tools call a session. Everything an agent does happens inside some run.

Each agent has its own list of runs, shown in the Runs tab.

How runs start​

Not every run starts with you typing a message:

  • Manual — you click New Run or send a message. The common case.
  • Scheduled — a schedule fires and the agent starts on its own. See Scheduled & Autonomous Agents.
  • Delegated — a parent agent hands work to a child agent, which creates a run on the child.
  • Subrun — the agent spawns a nested run inside itself to isolate a subtask (more below).
  • Recovery — Taurus resumes a run that infrastructure interrupted (a deploy, a container restart). You'll rarely notice these.

The run list marks each with a small icon: a clock for scheduled runs, a branch for subruns, a share icon for delegated ones.

Statuses​

At any moment a run is either active (running, paused for your input, waiting on a child...) or terminal (completed, stopped, errored). The full list with what-to-do notes is in Run Statuses.

The important thing: terminal runs are not dead ends. You can continue any of them.

Continuing vs. starting fresh​

Send a message to a completed or stopped run, and the agent picks up exactly where it left off — the whole conversation is still in its head. Start a New Run, and the agent begins with a blank slate: just its system prompt and memory.

Rule of thumb: same task → continue the run. Unrelated task → new run. Old history on an unrelated task isn't just useless, it actively gets in the way.

Forking​

Forking creates a new run from a chosen message: everything up to that point is copied in, everything after is discarded, and the original run stays untouched.

This is for "what if we'd gone the other way at step 5" moments. Fork from the message before the decision, give different instructions, and compare the branches. You fork from a message's dropdown menu — see Message actions.

Subruns​

An agent can delegate work to itself with the Subrun tool. The subrun gets a fresh, separate context window but keeps the same prompt, tools, and container. When it finishes, only its final report lands back in the parent run.

This is how agents keep their main context clean: a subrun can chew through fifty web searches or a noisy debugging session, and the parent sees just the conclusion. Subruns appear indented under their parent in the run list. Subruns can also run in the background, so the parent keeps working while they grind.

Keeping the list manageable​

Runs pile up. Two features help:

  • Labels — rename a run ("shipped auth fix") so you can find it later without scrolling by date.
  • Pinning — pin the runs you keep coming back to; they stay at the top of the list.

Further reading​