Routines

A routine is a task you save once and run again: on a schedule, such as every weekday at 9:00, or by hand whenever you want. Each routine has a name, instructions, a project folder, an optional model and a permission mode. Use routines for recurring work: a morning briefing on recent commits, a weekly dependency review, a draft of release notes.

AgenC Core keeps the routines and runs them. Each run starts a fresh session in the routine's folder, and you can open it like any other chat.

Open Routines

Click Routines in the sidebar, under New session. The page lists your routines, a box to draft a new one, and templates. On macOS and Linux you can also ask the agent in a chat to create or change a routine for you; see Create a routine from a chat.

Create a routine

There are three ways to start:

• Write what you want automated in the box ("What do you want automated?") and click Draft routine. The editor opens with your text as the instructions and its first line as the name. It does not read a schedule out of the sentence; set that in the editor.
• Click a template under "Or start from a template".
• Click New routine + for an empty editor.

The editor has these fields:

• Name (required).
• Description (optional): a short summary.
• Instructions (required): what AgenC should do on each run. Under the box you choose the model ("Core default" or a model of your default provider) and the project folder. New routines start in the default project folder from Settings; Change picks another.
• Schedule: see below.
• Behavior tab: the permission mode and the Enabled switch ("Paused routines do not start scheduled runs.").
• Notifications tab: Notify when a run finishes.

Click Create. Nothing runs or gets scheduled before that. Cancel with unsaved changes asks "Discard changes?"

New routines start enabled, in Read-only mode, with notifications on.

Templates

Five templates ship with the app. Each fills the editor; review the instructions and folder before you create it.

• Code briefing: "A concise summary of recent changes and what needs attention." Reviews the last day of commits, flags risky patterns and missing tests, and lists follow-ups. Weekdays at 09:00.
• Repository health: "Check the project's checks and config for gaps and mistakes." Reads CI workflows, scripts and configuration, and reports what is and is not wired into CI. Daily at 09:00.
• Dependency check: "Review declared dependencies, pins and lockfile drift." Mondays at 09:00.
• Release notes draft: "Turn recent commits into a readable release summary." Drafts grouped notes from the commits since the latest tag. Runs manually.
• Documentation check: "Spot stale instructions and gaps in project documentation." Mondays at 09:00.

The templates only read. They do not install anything, run the project's checks, publish or change files.

Schedules

Choose one of:

• Manual: "Runs only when you choose Run now."
• Hourly, at a minute you choose.
• Daily, Weekdays (Monday to Friday) or Weekly (on a day you choose), at a time you choose.
• Custom: a five-field cron expression, "minute · hour · day · month · weekday", with numbers, ranges, lists and steps.

Times use the time zone of the Core process, which the editor names. A routine never overlaps itself, and a change applies to future runs, not the one in progress.

When routines run

Routines run only while the computer is awake and Core is running. A run that was due while Core was offline is not run later. In practice:

• The Core that AgenC Desktop started stops when you quit the app, and with it the schedule.
• On Windows and Linux, closing the last window quits the app. On macOS the app keeps running after you close the window, until you quit it.
• Keep this computer awake in Settings > Connections does not cover routines. It only applies while connections are on.

Permission modes

A scheduled run has nobody to approve anything, so a routine's mode decides what it may do on its own:

• Read-only: "Reads this project and reports. Anything that would change it, reach the network or need your approval is refused."
• Approve edits automatically: "May edit files in this project without asking. Anything else that would need your approval is refused."
• Bypass permissions: "Runs without asking for approval. Its commands run in the sandbox and write only inside this project. MCP tools you have set up can still act outside the project and use the network without approval."

Plan mode is not offered, because nobody could approve the plan. Saving a routine in Bypass permissions also trusts its folder, as choosing that mode for a chat does. The routine list shows each routine's mode: "read-only", "auto edits" or "bypass". An older Core runs routines read-only only, and the field then shows "Read-only" without a choice.

Inside an open routine session, the permission and model chips do not change the routine. The app says "Change permissions in Routines for the next run." and "Choose the model in Routines for the next run."

Tools a run gets

A run uses Core's configured tools and plugins. When a Desktop window is open, the app also attaches its browser, terminal and window tools to the run. The run history says which: "Desktop tools attached.", or "Desktop tools unavailable" with the reason, such as "Desktop window is not open." On Windows the note in the editor says "Desktop controls for agents are unavailable on Windows. Routines can use Core tools."

A scheduled run can read routines but cannot create or change them.

Run, pause, edit and delete

Each routine row has a menu:

• Run now: starts a run. It works even when the routine is paused, and it is off while a run is active. The app says "Run started. Open History to follow its progress."
• Edit: reopens the editor. Save with Save changes.
• Pause or Resume: turns the schedule off or on. Pausing does not stop a run in progress.
• History: the run history.
• Delete: asks "Delete routine?" The routine and its run history are removed; session transcripts are kept.

The row itself shows the schedule ("Next:" with the date, "Runs manually", "Paused" or "No upcoming run"), the mode, and the last run's status.

Run history

Click a routine's name, or History, to see its last 50 runs. Each run shows its status (Starting, Running, Needs permission, Completed, Failed, Cancelled or Interrupted), when it started, whether it was manual or scheduled, when it finished, the Desktop tools note and any error.

• Open session opens the run's session. While the run is active you can watch it but not type. After it ends, the session is an ordinary chat and you can ask follow-up questions.
• Review request appears when a run waits for permission.
• Cancel run stops an active run. If Core cannot prove that the run stopped, it blocks new runs and deletion until it can.

A run that Core's restart interrupted ends as Interrupted and is not repeated. A scheduled run's session appears in the sidebar once you open it, named after the routine.

Notifications

With Notify when a run finishes on, the app shows a desktop notification titled "AgenC · Routines" when a run finishes, fails or is interrupted, for example "Morning briefing: finished." Click it to open the run's session.

The notification needs Desktop notifications on in Settings > General and appears only while the app window is in the background. If the app-wide switch is off, the routine editor says so.

Create a routine from a chat

On macOS and Linux you can ask the agent in any chat to create, change, run, pause or delete a routine, for example "create a routine that checks the dependencies every Monday at 9". The agent uses the same routines you see on the page. On Windows the agent does not get these tools.

A routine the agent creates gets the chat's permission mode, or a narrower one if you ask. It never gets a wider one. A chat can also not change the instructions, folder, model or mode of a routine that has wider permissions than the chat; it can still pause, rename or reschedule it. A chat in plan mode, or in "Plan without changes", can look at routines but not change or run them. Change anything else on the Routines page, which can set any mode.

The routine's folder defaults to the chat's folder. Another folder must be one of your Desktop projects.

Limits

• Up to 100 routines, and 50 runs of history per routine.
• A routine's instructions can be up to 16 KB.
• On Windows the agent cannot drive the app's browser or terminal during a run.