Agent settings
Inspect › Settings (4, then the Settings tab; /inspect settings) is Studio's agent config as rows.
On this page
Inspect › Settings (4, then the Settings tab; /inspect settings) is
Studio’s agent config as rows. Instructions, Tools and Compaction open
dialogs; Autonomous, Autostart, Receive messages, Host access, Visibility,
Send mode, Inbox mode and “New MCP tools need approval” flip in place with
Enter or Space. Turning on autonomy or host access, and LAN or public
visibility, asks first with Studio’s warnings. l locks the highlighted
section so the agent cannot change it itself (locked_fields; you still
can). Every change re-reads the config, changes only what you edited, saves
and reports the result.
Instructions
/instructions opens an editor in the app: Enter is a newline, Ctrl+S
saves, Ctrl+O continues in your external editor, Esc cancels (asking when
there are unsaved changes). /instructions edit (or e on the Settings tab)
goes straight to the external editor. If the daemon’s copy changed meanwhile,
saving asks before overwriting.
Tools
/tools [filter] lists Studio’s tool groups plus one group per MCP server,
each tool with enabled, shown, approval required and locked:
| Key | Action |
|---|---|
Space, Enter | Enable / disable (on a group header: the whole group, locked tools skipped) |
v | Show / hide in the agent’s tool list (hidden tools stay callable from code) |
r | Require approval: a model call waits for you |
l | Lock / unlock a built-in tool for the agent |
/ | Filter |
Studio’s rules apply: a locked tool cannot be enabled or shown, turn tools
(say, ask, sys_set_state) have no approval toggle, and a hidden tool
must be enabled. Refusals show as warnings.
Compaction
When a loop’s context reaches its threshold, its history is summarized
(compacted). /compaction lists main and every inner loop with its threshold
(default, inherits main, or set) and the context in use:
| Key | Action |
|---|---|
Enter, e | Set the threshold: 80000, 80k, default, inherit |
d | Back to the default (main) or to main’s threshold (an inner loop) |
c | Compact that loop now (asks) |
/compaction 80k sets the selected loop’s threshold directly. /compact
compacts the selected loop now. What fills the context: Context.
Model
/model picks the selected agent › loop’s model (a picker, or
/model [provider/]model). On main it changes the agent config; on an inner
loop it sets that loop’s override (/model inherit clears it). See
Models and providers.
Config
Inspect › Config (/config) shows the full agent config as a tree, or JSON
with /json on. e (or /config edit) opens it in your external editor;
on save it is validated against the schema and the changed keys are shown for
you to confirm.
One-shot: adf config <agent> prints it.
Other Inspect tabs
Status (state, loops, triggers, website), Usage (tokens by model), MCP,
Channels, Identities (metadata only), Logs (f follows), Tables (the agent’s
database tables), Tasks (/tasks: approvals and async tool calls; y, n,
a, N feedback, A approve all) and Events (this agent’s live events; l
narrows to the selected loop). Secrets are redacted everywhere.