Interaction Modes (Human-in-the-Loop)

Arbor runs fully autonomously by default, but you decide how much oversight you want. The interaction mode controls when — if ever — the agent pauses to consult you, and a set of live controls let you steer a run as it happens.

The four modes

Mode	Behaviour
auto	Fully autonomous. The agent never pauses for input.
direction	The agent asks where to explore at key junctions.
review	The agent asks you to approve or edit ideas before running them.
collaborative	Both the direction and review gates are active.

Setting the mode

Per run, on the command line (alias --mode):

arbor --mode review
arbor --interaction-mode collaborative

Or durably in config:

research_config.yaml

ui:
  interaction_mode: auto      # auto | direction | review | collaborative

As always, the CLI flag overrides the config value. See Configuration → When settings disagree: precedence.

What a gate looks like

When a gate triggers, the agent pauses and asks for your input — in the terminal dashboard and, for interactive runs, in the Web UI. In review mode, for example, you can approve a proposed idea as-is, edit it, or redirect; in direction mode you nudge which part of the tree to expand next.

Unattended runs still make progress

Pass --no-dashboard-input and gates auto-continue after a timeout instead of blocking forever. This lets you run a review/collaborative study unattended: it pauses briefly for input, then proceeds on its own if you're away.

Steering a run live

Independent of the mode, you can always influence an active run from the terminal dashboard (and the interactive Web UI) using slash commands:

Command	Use
/steer <message>	Inject guidance directly into the research agent.
/ask <question>	Ask the read-only companion about the run (doesn't change it).
/skill <name...>	Ask the agent to load a Skill on demand.
/pause / /resume	Pause after the current step, then resume.
/tree, /evidence, /branches	Inspect state before deciding how to steer.
/abort	Stop the run.

Choosing a mode

You want…	Use
Maximum autonomy / benchmarks	auto
To keep the agent on a research direction you care about	direction
A check on each hypothesis before compute is spent	review
Close collaboration on a hard problem	collaborative
Hands-off but with light supervision	any mode + --no-dashboard-input

For how gates fit into the search loop, see How It Works → Human-in-the-loop.