Chapter 13 : Delegating: permissions and background tasks
When you ask a question, Jay answers and stops there. When you hand it a goal, it picks the steps and tools itself, looks at the result, corrects course, and keeps going until the deliverable is done. That is what makes it an agent. Knowing how it runs lets you choose what you hand over, and how far.
The agent loop
- Your brief
- 1 · Plan
- 2 · Act
- Permission?
- 3 · Observe
- Your brief + the folder, the JAY.md
- Permission? granted: the tool runs · denied: Jay knows and adapts
On each turn, Jay breaks the work into steps and picks a tool, and this is where your access level decides: the tool runs on its own, or a card asks for your approval. Jay then reads the result, adjusts its plan, and starts again until the "done" criteria are met. A task list appears under its reply as it goes, and an end block closes the turn: the summary, its question if it has one, the files produced, the time and the tokens.
The three access levels
You choose the level in the bar, on the access selector right after the project, per session. It doesn't change what Jay can do: it changes what it does without asking you. Settings, the bar and this guide use the same three words.
| Level | What Jay does | When to use it |
|---|---|---|
| Ask | Asks for confirmation before every action: writing, a command, a connector, a click on a page. Reads always go through. | To get started, and for any sensitive task |
| Auto | Approves everyday actions on its own: commands, calendar, reading and sorting email, drafts, Drive, memory. Shows the diff before any write to a file, and asks before acting on a web page. | Tasks you know well |
| Full | No confirmation, except for the few always-confirmed actions below. Jay writes, runs its commands and clicks on pages without asking. | Well-honed tasks, in a folder where nothing is irreversible |
One access level per position. A session keeps two levels, one for Work and one for Code, each on its own selector. Work starts in Full: you write documents and Atelier pieces there, and Jay doesn't interrupt you at every paragraph. Code starts at the level you chose in Settings (Ask out of the box). Switching from one to the other applies the level of the position you land on, and a message warns you when access widens. If you'd rather Jay also ask in Work, set "Default access in Work" in Settings, AI behavior.
When a permission card appears, the bar becomes the question: Enter accepts, Esc declines, and a command's card says what it touches and what it risks. When you approve often, Jay tells you: "Approving a lot? Auto mode lets everyday actions go through." Switch when you're comfortable, not before.
What stays confirmed, even in Full. Anything that goes out to a person: an email, a Slack message or one sent from the Mac, a Notion comment, a sent invoice, an event whose guests get notified, a message to another session, and soon a LinkedIn post. Sharing a Drive file too. And a command you've denied in Settings stays denied, even hidden inside a chain.
Plan mode. In the Code position, the "+" offers Plan, or type /plan anywhere: Jay reads and analyzes, but writes nothing; a read command goes through, any other command is asked. It presents a plan, you approve it or have it regenerated, then /build carries it out. It's the right reflex before any task that touches a lot of files.
The delegation brief
Say you ask Jay to "find suppliers for the event." It will search, compare, maybe draft emails, and hand you something very different from what you had in mind. An agent is only as good as its brief. The more autonomous it is, the more the brief must spell out the deliverable, the limits and the moment it's done.
Goal: [the end result, in one sentence] Deliverable: [file, format, location in the folder] Context: [what Jay needs to know, and where to find it in the folder] Resources: [folders, connectors, sites to use] Limits: [what it must not do: send, delete, buy, publish, modify a given folder] Done when: [the verifiable criteria that say it's finished] Before you start, show me your plan in 5 steps max and wait for my approval.
Background tasks and sub-agents
A long task doesn't have to block your session. Say "do it as a background task": Jay launches a separate sub-agent, with its own context and the same tools, and you keep working. The result comes back to the session on its own, even if you've moved on to something else in the meantime. Up to three background tasks run at the same time. The Activity tab of the drawer shows what's in progress, and the header says how many agents are working.
Each sub-agent gets a role, and the mode that goes with it: explore (read and search, fast), build (write), review (check a piece of work), or expert (the deepest reasoning). It inherits the session's context and returns a report suited to its role. A sub-agent works on its own, without asking you anything: with Full access it acts alone, otherwise it stays limited to what your level lets through without a card. Give it a tight brief, with clear limits, and keep it for tasks where a mistake can be fixed. For longer parallel work, Jay can also suggest a sub-session, attached to yours, that you follow in the rail. And in Customize, Advanced, Sub-agents, you create specialized agents with their own brief, which Jay calls when the task fits them.
A long turn no longer stops on its own after a fixed time: Jay goes all the way to the deliverable, and the Stop button stays yours.
The terminal
In the Code position, the Terminal button in the bar opens the Terminal tab of the drawer: a real terminal in the project folder, shared by all the project's sessions. Jay runs its commands there and reads their output; you can type your own. When it gives you a command to run yourself, it writes it in a block you copy with one click. The commands you allow or deny by default are set in Settings, AI behavior.
When not to delegate
- When a question is enough. A full task for a one-line answer is wasted volume.
- When a mistake is costly and can't be undone. Payments, mass mailings, deletions: keep the final decision, with Ask access.
- When you don't know yourself what you want. Clarify first in conversation, or with /goal. Delegate afterward.
Ten tasks to hand to Jay
| Task | What you ask |
|---|---|
| Tidy a folder | "Sort my Downloads folder by type and by project, rename the files consistently, and list what looks useless without deleting anything." |
| Expense reports | "Read the receipts in this folder and fill in the Excel expense report template." |
| Monthly report | "Using the files in this folder, write this month's activity report in the same format as last month's." |
| Monitoring | "Monitor my 5 competitors this month and summarize the changes in one page." |
| Presentation | "Turn this report into a 10-slide presentation for the committee, using the brand kit." |
| Contact database | "Merge these three contact files into one, with no duplicates, and flag incomplete records." |
| Meeting prep | "Prepare a brief for each of my meetings this week, from the calendar, email and the folder." |
| Document summary | "Read the 20 PDFs in this folder and make a summary table with the key points of each." |
| Reminders | "Find the unpaid invoices in Pennylane and prepare a draft reminder for each client." |
| Comparison | "Compare these 4 supplier quotes in a table and recommend the best one based on my criteria." |