How to give the AI a standing role you reuse
Short answer: Save it as an Agent: a name, a standing prompt, a model, and a tool allow-list, configured once in the Workers panel and run on demand from then on. Because an agent never sees your chat history, its prompt has to be self-contained: whatever background or constraints the job needs go into it directly.
Steps
- Open the Workers panel (Ctrl+4) and switch to the Agents tab.
- Type a name into the “New agent” field and press Enter, or click the + button, to open the agent editor. (Or, from a chat that’s already doing the job, ask the AI to “save this as an agent that does X”; it opens the same editor for you to review before anything is saved.)
- On the Basics tab, pick a model and write the standing prompt into the Instructions box.
- On the Tools tab, check off exactly the tools the agent may call; anything unchecked simply isn’t available to it, even if the same model would have access to it in a normal chat. Add a validator model here too if you want a second model reviewing specific tool calls before they run. For tools that send things out (email, Discord, Notion), this tab also holds each tool’s Auto-send destinations list: the destinations the agent may send to without pausing on an unattended run.
- On the Behavior tab, set an input folder, output folder, deliverable file, or seed notes if the job needs them, then save.
- Run it any time from its card in the Agents tab (the Run button, or Run now in the card’s menu), or delegate to it from chat once it’s saved: inline (the chat waits for the result) or in the background (the chat keeps going, and the result posts back into the conversation when the agent finishes).
Notes & gotchas
- The tool allow-list is absolute: an unchecked tool can’t be called by that agent under any circumstance.
- An agent doesn’t schedule itself: its editor has no schedule field. To run it on a timer, or fire it from an incoming email or Discord message, wrap it in a one-step workflow (step kind “Run a saved agent”) and attach a trigger there.
- Anything the agent sends out (email, Discord, Notion) only goes to destinations you pre-authorized in its Auto-send destinations list on the Tools tab. Anything not on that list parks the run for your review instead of sending while you’re away.
- Every run gets a card in the Workers panel’s History tab with one-click Revert this run, so a bad run’s file changes are easy to undo.
- Agents can’t delegate to other agents; that stays chat-only. They can generate images, speech, and music (held to a daily budget cap, default 50/day); video generation is chat-only.