Jobs & Trust Modes

When the agent needs to do real work, like running a command or a long task, it creates a job. Trust mode decides whether you approve each job first.

Trust modes

Mode	Command	Behavior
Supervised (default)	`/mode 1`	The agent queues each job and waits for you to approve or reject it.
Auto	`/mode 2`	The agent runs jobs as soon as it creates them.

To set a permanent default, edit ## Agent Trust Level in BRAIN.md (supervised or auto).

// BE DELIBERATE WITH AUTO MODE In auto mode the agent can run commands, change files and act on connected accounts without asking. Use supervised mode for workspaces with access to sensitive data or systems, and switch to auto only for work you're comfortable delegating.

Managing jobs

Open the Jobs page for a workspace (/workspaces/:id/jobs) to see every job and its status, or use chat commands:

/jobs                  list all jobs
/jobs approve <id>     approve one job        /jobs approve all
/jobs reject <id>      reject one job         /jobs reject all
/jobs run <id>         run a job now          /jobs run all
/jobs halt             stop all running and queued jobs
/jobs clear            remove completed and failed jobs

Job states: pending, running, completed, failed, or paused waiting for your input when the agent needs an answer to continue.

When a job fails

## On Job Failure in BRAIN.md controls what happens next. The default is continue. The agent can also retry failed model calls automatically (Global Settings → Retry on failure).

Job timeout

Each job has a wall-clock limit, 15 minutes by default (jobTimeoutMinutes in ~/.agentos/server.json). A job still running past the limit is stopped and marked failed, so nothing hangs silently in the background. Set the limit to 0 to turn it off.

Approving installs

Package installs (such as npm install or /install plans) always go through approval unless you turn on Auto-approve installs globally or set ## Auto Approve Installs in a workspace's BRAIN.md.

Queuing jobs from outside

Other programs can queue or run jobs through the webhook API. See Agent Interfaces & API.