Agents, Templates and Triggers

FAC Chat agents (workflows) run on FAC Cloud and act on your site through your MCP tools. This page covers what you see in FAC Chat and what your site enforces. Building and editing agents needs the System Manager role; other users can view agents and their runs.

Who an agent runs as

Every agent has a runtime user ("Runs as"). A run uses that user's MCP tools and approval settings, not those of whoever opens the builder. New agents run as the person who created them.

• The builder's tool list and the green/red tool badges show the runtime user's tools.
• Run this node runs one step as the runtime user, so a test uses the same tools as the real run.
• Checking another user's tools is limited to System Managers.

The Setup checklist

The Setup panel in the builder lists what an agent needs to run on its own: how it starts (schedule or event triggers), who it runs as, whether its write tools are approved, and which model it uses. Each row shows its current state and a shortcut to fix it (Schedule, Triggers, Settings, Re-check).

Write actions in unattended runs

Scheduled and triggered runs have nobody to click Approve. A write tool (for example create_document) only runs unattended when the runtime user has set it to Always allow. Otherwise the step is skipped and the run still ends Completed. The run shows Completed · N actions skipped, and the run details list the Skipped actions.

When you activate an agent whose write tools are not set to Always allow, the builder warns that they won't run unattended and names the tools, so you can fix them first. The warning appears only when FAC Cloud can report the tool's approval setting; if it cannot, no warning is shown.

Templates

Importing a template asks for its settings. Each setting has a type: a link to a record on your site, an email, a number, a checkbox, a choice, or text. Before anything is created, FAC checks the values on your site: required settings must be filled in, numbers must be numbers, and a linked record (a Company, a Warehouse) must exist here. Every problem is listed at once. Link and email settings are never pre-filled with sample values. Importing an agent template needs System Manager.

A template that starts on a document event also brings its trigger, and a template that runs on a schedule brings its schedule. Both arrive switched off: the Setup checklist's "How it starts" step tells you which one to turn on, and nothing runs until you do. If your site cannot create the suggested trigger (for example, the DocType is not installed), the agent still imports and you add a trigger yourself.

Triggers

A trigger starts an agent when a document is created, updated, submitted, cancelled or deleted.

• Filters narrow which documents fire it. Editing a trigger (even only its title) keeps its filters.
• Test checks the trigger against the latest document you can read and shows the data the agent would receive. If a filter would stop it, the result names the failed filter. The test checks filters only, not the event or which fields changed.
• Loop guard. An agent that writes back to the document that triggered it can loop. One trigger fires at most 5 times per document per minute and 20 times per document per hour; further fires are refused and logged once. These site config keys change the limits, and 0 turns a limit off: fac_trigger_breaker_max_fires, fac_trigger_breaker_window_seconds and fac_trigger_breaker_max_fires_per_hour.
• What the agent receives. The document with Password fields removed, including in child tables. Above 256 KB the largest child tables are left out first (listed in dropped_child_tables); only a document that is too large on its own is reduced to its name and title.

Schedules

A schedule has a cron expression and a time zone, so "09:00 every weekday" means 09:00 in the zone you pick. The zone defaults to UTC; your browser's zone is offered in the list.

Run history

The run history lists each run with its status, and Load more fetches older runs. Opening a run shows each step's input and output; very long values are shown as their first 10,000 characters, and that partial view has no Download. A run's final output can be copied, downloaded, or opened Full screen.

Data tools inside agents

Agents use the same tools as chat. To keep a run's context small, generate_report returns at most 500 rows unless asked for more (up to 5000), run_database_query returns at most 1000 rows, and list_documents pages with offset / next_offset. See the Tool Reference.