Agent settings
Agent settings
Open Settings ▸ Agents, then select an agent to open its editor. The editor has tabs for Personality, Model, Skills, Secrets, and Usage (a Network tab also appears when the residential-proxy feature is enabled). Changes save per tab.
Personality
The Personality tab holds the agent's display name and its soul.md — plain-text rules
that shape how the agent thinks, writes, and behaves. Write it like a job description.
Model
The Model tab sets this agent's LLM. You pick a provider (Anthropic is recommended), then a model, and supply an API key for that provider. Saving redeploys the agent's service with the new settings.
- Models are listed newest-first. Anthropic leads with Claude Fable 5
(
anthropic/claude-fable-5, native 1M context), followed by Opus 4.8 / 4.7 / 4.6 / 4.5, the Sonnet family — where Claude Sonnet 5 (native 1M context) is the default selection — and the Haiku family. - Changing only the model reuses the stored key — you don't need to re-paste it. Switching provider requires a fresh key, since a key for one provider won't authenticate against another.
- Reset to Default clears the custom provider and key. The agent falls back to the managed OpenRouter default (Kimi K2.5); workspaces created before the managed-key rollout reuse their provisioning key.
The same per-agent provider / model / API-key controls are also available on the workspace Settings page under the LLM tab — pick the agent from the selector there. For the full provider catalog, bring-your-own-key details, and how managed credits work, see LLM configuration.
Timezone
Timezone isn't a separate tab — it's a card at the bottom of the Model tab. Each agent runs
its machine in a specific timezone (an IANA name such as Asia/Kolkata). The timezone drives
how the agent interprets "today" and sets the default for new recurring tasks.
Saving a timezone restarts the agent.
Skills, secrets, and network
The Skills tab toggles which installed skills this agent may use; each enabled skill shows a sync status (Pending / Syncing / Synced / Failed), with a Retry on failure. The Secrets tab grants the agent the credentials it needs. The Usage tab shows the agent's machine runtime over the trailing 30 days — a live-session indicator, a session count, average and longest run, and a dated activity log.
When the residential-proxy feature is enabled, a Network tab lets you route the agent's browser traffic through a residential IP so sites are less likely to block it; toggling it restarts the agent. See Skills and Secrets for details.
Lifecycle
When a workspace is provisioning, restarting, or recovering, its agents pass through initializing before returning to running. A running agent is online and able to handle chats, tasks, and messages from your connected integrations. Saving a model or timezone change briefly redeploys (or restarts) the agent.
Deleting an agent
In the agent editor, open the ⋯ menu and choose Remove agent. Removing an agent stops it right away and drops its cost from your next billing cycle; its chats and files stay in the workspace. Every workspace needs at least one agent, so the last remaining agent can't be removed — add another first.
