Agents & Automation
Create a saved Agent
- Open the Agent tab and choose New agent.
- Choose one of 27 specialist templates or Custom, then edit the name, specialty, and instructions describing the work it should own. Templates can fill the mode, permissions, and a recommended connected model; review these settings before saving. Undo restores draft settings replaced by a template.
- Choose its provider account and model, then review permissions, tools, and connected services. Its chat starts in the current project.
- Save the Agent. Its first chat opens so you can start working immediately.
You do not need a schedule or event to create an Agent. A saved Agent is a reusable identity with instructions, a model route, and access settings. It can own several conversations and have optional automation triggers.
Your companion offers another way to create or explicitly link that same saved identity, starting with an offline character and name. It can save before a model is configured. Changing the character later does not reset the agent's conversations, permissions, instructions, or memory, and setup does not activate automation.
For example, create a “Project reviewer” that checks changes and reports missing evidence. Chat with it first; add a daily schedule when you want the same review to recur.
Chats, configuration, and activity
Use search and filters to find a saved Agent. Select it to return to its conversation, or open Manage Agent to inspect configuration, chats, triggers, and recent activity. New chat starts another conversation using that Agent’s profile.
Work and Agent remember separate selected conversations and drafts. Switching tabs restores the appropriate chat without moving its draft into the other workspace. A conversation retains its Agent association when reopened.
| Surface | What it shows |
|---|---|
| Agent | Saved Agent collection and the selected Agent’s conversations |
| Manage Agent | Instructions, model, access, chats, triggers, and recent activity |
| Activity | Schedule and event records, execution outcomes, and supported retry actions |
| Connections | Shared event sources, health, and their consumers |
| Runtimes | Independent services and SSH deployment settings |
| Overview / Runs | The selected conversation’s progress and execution records |
| Attention | Questions, permissions, approvals, recoverable runs, and configuration problems requiring a decision |
Use the full workflow in an Agent chat
Saved Agent conversations support ordinary chat, Work, Plan, Grill, Goal, Duo, Task Capsules, and team workflows. Start with the composer’s workflow menu and choose the approach that fits the task.
Ordinary turns use the Agent’s saved model route, instructions, and access. Duo, Capsule, and team stages use the profiles and models explicitly chosen for those stages, while the conversation keeps its saved Agent association. Each stage remains subject to its profile limits and the chat’s permission rules.
Duo pairs a planner and builder in the conversation. Persistent Goals continue an objective across turns. Task Capsules retain approved plans, verification, and recovery. Agent Teams coordinate named roles.
Use helpers and configured teams
A saved Agent’s Work conversation can delegate bounded helper work through the existing collaboration tools. Helpers return findings and evidence to the coordinating agent, which remains responsible for the combined result. The active workflow, permissions, and task limits govern that delegation.
To collaborate through named saved profiles, select a configured team. Its dispatcher and specialists use their configured models, access ceilings, and budgets. A saved profile is reusable across chats and teams; a helper is scoped to its coordinating task. See Agent Teams.
Add triggers when the work repeats
In Manage Agent, choose On a schedule, When an event arrives, or When a price changes. Configure the timing or matching conditions, inspect the prompt and execution environment, and save the trigger. A saved Agent can have multiple attached triggers; each trigger has its own configuration and occurrence history.
Incoming events can use configured Gmail, Telegram, or signed webhook sources. Price conditions use a configured source, symbol, and threshold. Incoming-event triggers can start paused so their configuration can be reviewed first.
A source connection for receiving an event is separate from permission to act on a service. Review the Agent’s allowed action connections explicitly. Clearing all action grants keeps the selection empty; webhook and price-feed identifiers do not grant action access.
New saved Agents start with the Read only access level. Choose Workspace edits or Computer control when the Agent needs those capabilities, and review its tools and connected services. This profile ceiling is separate from the chat’s Ask every time, Accept file edits, or Full access approval mode: Full access does not override a read-only profile. See Permission Modes.
For a new Agent, choosing a write-capable access level enables Allow all services and extensions when no service policy has been set. That option includes services connected later; turn off individual services or choose explicit access as needed. Chat permissions still apply.
Coordinate on the Board and Calendar
Assign board cards, calendar events, and notes to saved Agents. A board handoff creates a draft chat in that board's workspace. Completed work offers previews, review, and revision requests. A Board can open in a separate window and stays synchronized with the main workspace.
Agents can read the workspace Board and create, move, edit, comment on, or delete cards; each change is attributed to the Agent that made it. Read-only Agents receive only board reading. Creating, editing, and commenting on cards ask in Ask every time and run automatically in Accept file edits; deleting a card asks in both. The Agent's Workspace files and Edit workspace files toggles gate reading and changing the board.
With Calendar access allowed in macOS, Agents can list events. Creating, updating, and deleting events follow the normal write permission decision; read-only Agents can only list.
Workflows and approvals
Automation workflows sequence Agent, Condition, and Approval steps. The simulator previews templates, branches, outputs, and approval cards without model calls or chat history. A chat’s workflow menu is separate from this automation step editor.
Each occurrence records attempts and approval state. After interruption, inspect the durable record before resuming. An uncertain external action is not silently repeated. Attention holds the decisions; routine successful activity remains in Activity.
Pause, run, and recover
Use Pause to stop a trigger’s automatic starts and Run now for supported manual execution. Pausing a trigger is separate from stopping an active chat. Receipt, delivery, and execution outcomes are distinct: a previous successful chat does not make a later failed delivery successful.
Desktop work needs Locus and an available Mac. With the opt-in independent runtime enabled in direct-download, wallet-free Locus, supported work can continue after the app closes while its host remains available. Native desktop actions wait for Locus. A hosted model account alone does not provide a background host. See SSH & Remote Runtimes.
Review connection errors, unavailable accounts, and interrupted actions before retrying. Gmail polling fixes in 3.0 improve event delivery, but a disconnected source still needs attention.
Your saved Agents, conversations, permissions, and queued work remain in Locus when an optional plugin window closes or fails.