# Agent defaults and external runners

Open **Settings → Agent defaults** for the baseline inherited by new sessions, helper limits, and the state of external agent runners.

## Set the baseline
Choose the default agent and approval mode. Review sandbox settings where an ACP provider uses them. The persistent **YOLO mode** setting skips tool-approval confirmations and remains enabled until turned off, unlike a chat control with an expiry. Use it only with access and execution boundaries you intend.

## Bound helper work
Set maximum concurrent subagents, turns, timeout, minimum free memory, and allowed working directories. Enable orchestration only when you want the agent to delegate. A resource limit can stop a helper; it does not mean its task finished.

## Inspect runners
Runners are external agent CLIs the installation can drive. Re-check them to get a current measured health reading. An overdue reading is stale, not a new health probe. Runtime agent identifiers and capabilities must match the runner actually installed.

Unattended runs can require a verified adapter. A fetched-at-launch adapter, changed adapter, or missing provenance can block an unattended spawn even when an interactive chat is allowed.

## Session lifecycle
Concurrent ACP sessions depend on backend support. Idle release frees a runner hold without deleting the chat. Durable worker sessions require tmux; when supported they can make work resumable after a gateway restart. Reconnect and inspect the run before starting a duplicate.

## Optional self-QA
The self-QA companion watches a configured repository for user-impacting commits, runs scenarios, and files Inbox items and Tasks for failures. It is off by default and consumes execution resources. Proposed fix branches are for review; they are not automatically merged or pushed. Enable only with a deliberate repository and scenario limit.