Your First Agent Session¶
Choosing Your Interface explains why agents change what artificial intelligence (AI) can do for you; this page has you run one, once, on a folder that cannot be hurt. By the end you will have watched a model read files, ask your permission, run a tool, and hand back work with evidence you can check, and you will know which settings are worth changing, and when.
Before You Start¶
- Install an agent: the Claude Code desktop app or the ChatGPT desktop app (Codex is a mode inside it). Codex is included with every ChatGPT plan, including Free; Claude Code needs a paid Claude plan (Pro or above), so take the ChatGPT route if you have no subscription.
- Optional but worthwhile: the local toolkit, so document conversion works when you get ambitious later.
The Walkthrough¶
Twenty minutes, five beats
A folder
of copies, so no mistake matters
One real task
a deliverable, not a question
The prompts
read each one; you are the gate
The loop
read, write, check, continue
You review the output, yourself
Green: your safety net. Amber: your control. Filled: your judgment.
- Make a practice folder of copies. Create a folder on your desktop and copy four or five course readings or papers into it (PDFs are perfect). Copies, not originals: your first session should be one where no mistake matters.
- Open the agent in that folder. In Claude Code, open the folder; in the ChatGPT app, switch to Codex and open the folder as a local project. That folder is now the agent's working area: by default it works there and asks before going beyond it. Before you send the task, check the permission mode: in Claude Code choose Manual (the mode selector in the desktop app, Shift+Tab in the terminal), which asks before each edit; in the ChatGPT app choose Ask for approval, its most cautious mode, which asks before the agent goes beyond the folder. Claude Code now usually starts in Auto, which hands approvals to a reviewer and would hide the prompts this walkthrough is about; What Each Agent Calls the Rungs maps each agent's modes onto the ladder.
-
Give it a real task. Paste this, or your own version:
Read the documents in this folder. Create a file called summary-table.md with one row per document: title, source and year, a five-sentence summary, and the three points most worth teaching from it. Tell me which document was hardest to read and why.
-
Watch the permission prompts, and read them. In Claude Code's Manual mode, reading files inside the folder you opened generally will not prompt, but the moment the agent wants to write the new file, it asks. Codex's Ask for approval mode gives the agent more room: it edits inside the folder without asking and prompts before it uses the internet or changes anything outside the folder, so you may see no prompt at this step. This is the habit to build on day one: the prompt tells you exactly what the agent wants to do, and you are the gate. Approve what matches your request; deny anything that surprises you, and ask the agent why it wanted it.
- Watch the loop. The agent reads each file (tool call), builds the table (generation), writes the file (tool call), and typically re-reads its own output to check it (tool call again). That loop (act, check, continue) is what separates an agent from a chat window.
- Iterate like it is a conversation, because it is. "Add a column rating each paper's difficulty for second-year students." The agent edits the file it already made; nothing is regenerated from scratch.
- Review the work. Open summary-table.md yourself. Agents make verification cheap, and the habit of looking at what actually changed is what makes them safe.
If the agent says a tool is missing, it will name exactly what to install; that is normal, and it is how your machine gets equipped over time. Newly installed tools need the app restarted before the agent can see them.
The Settings That Matter¶
Every agent has settings worth a few minutes of your attention, and five matter most. Their names differ from one company to the next and change often, so this section says what each one does and where to start. Each company page gives the exact names its agents use, with worked examples of the commands that change them: ChatGPT (OpenAI), Claude (Anthropic), Gemini (Google), and Grok (xAI).
- Permission mode is the big one: how much the agent does before it asks you. This site describes the modes as rungs of a ladder (the figure below). Ask first asks before consequential actions, such as editing files, running commands, or reaching the network: right for your first sessions and for unfamiliar folders. Edits flow lets changes to files in the working folder through without asking (in some agents that includes deleting files there), once you trust the workflow: right for repetitive editing sessions. Checked autonomy hands approval to an automatic reviewer, a second AI model, which approves requests in your place and surfaces only the ones it flags: far fewer prompts, but the reviewer can be wrong (Anthropic and OpenAI both say so), so it suits work you would approve anyway. No prompts skips asking entirely; Anthropic and OpenAI recommend it only for an isolated environment such as a container or a virtual machine, because a mistaken or manipulated command, such as deleting files or sending data out, runs before you can see it. Plan has the agent read and propose without changing anything: the look-before-touching mode, ideal for "tell me what you would do," and useful at every rung.
- Effort sets how hard the model reasons before it answers. Drop it for mechanical batch work (renaming, reformatting, applying a known fix everywhere): the work gets faster and cheaper with no quality loss where no judgment is needed. Raise it only for the genuinely hard step. Two of the companies advise against living at the top: Anthropic's documentation says its highest level "may show diminishing returns and is prone to overthinking," and OpenAI's says plainly of its two highest levels, "Most tasks do not need Max or Ultra."
- Orchestration splits a large task across helpers working in parallel (subagents) and combines what they return. The trade-off runs both ways: markedly more thorough on large work that splits into independent parts (audits, sweeps, many-file changes), and markedly more time and usage. Turning it on for the big task, not for the afternoon, keeps that spend where it pays.
- Fast mode, which some agents offer at a higher price, is worth it for back-and-forth where you are waiting on each reply, and wrong for long tasks you are not watching.
- A folder brief, a text file of standing rules the agent reads before any work, makes your preferences permanent. See Standing Setups.
The Starting Posture¶
This site's recommendation, stated as a table so you can disagree with it precisely:
| Setting | Start with | Graduate to | Rarely worth it |
|---|---|---|---|
| Permissions | Ask first, with Plan at any stage for a look before touching | Edits flow or Checked autonomy, once a workflow has earned trust | No prompts, outside an isolated machine |
| Effort | The default | Lower for mechanical batches; higher for the genuinely hard step | The highest level as an always-on default |
| Orchestration | Off | On for large tasks that split into independent parts, accepting the time and usage cost | On for routine questions |
| Folder access | One task-specific folder | Additional folders added deliberately | Your whole home directory or disk |
Permissions: a ladder you climb per workflow, not a dial you set once
Plan mode: read and propose, change nothing; useful at every rung.
Ask first
asks before edits, commands, and network access
Edits flow
file edits go through, for a workflow that has earned it
Checked autonomy
a reviewer surfaces only what it flags, and it can be wrong: keep reading the surprises
No prompts
nothing asks at all: built for isolated machines
A new kind of task is a good moment to step back to ask first, because the workflow has not yet shown what it will do.
What Each Agent Calls the Rungs¶
Each company names its modes in its own way and draws the boundaries a little differently, so treat the rungs as approximate. The same name can sit on different rungs (a mode called Auto lets edits flow in one agent and hands approvals to a reviewer in another), and an app's most cautious mode may still let edits inside the folder through without asking. Before you rely on a mode, read its own description on its company page: each agent's name in the table links there, and the same page's commands and shortcuts show how to switch. In the table, CLI means the command-line interface, the version that runs in a terminal.
| Rung | Codex (OpenAI) | Claude Code (Anthropic) | Antigravity CLI (Google) | Grok Build (xAI) |
|---|---|---|---|---|
| Ask first | Read Only (CLI) | Manual | default | Ask |
| Edits flow | Auto (CLI); Ask for approval (desktop app and editors) | Accept edits | accept-edits | acceptEdits (launch option) |
| Checked autonomy | Approve for me (Auto-review in the CLI) | Auto | Not offered | Auto |
| No prompts, for isolated machines | Full access | Bypass permissions | always-proceed (setting) | Always-approve |
| Plan, at any rung | Plan mode (/plan) |
Plan | plan | Plan |
Habits Worth Keeping¶
The folder you open is your main control over what the agent sees. Opening only the folder a task needs keeps the agent focused, makes its searches faster, and means you know everything it could read. Because the agent may read any file in that folder, everything in it is effectively shared with the model, not only the files the task is about; if a folder also holds something like a class roster, copy the files the task needs into a fresh folder instead. For a new kind of task, working on copies first is still worth the minute it takes: you see what the agent does before anything that matters is touched. The AI Agents guide covers prompt injection and the rest of what is worth knowing once you are running sessions routinely.
Next: Standing Setups, so the assistant keeps your context between sessions.
Content last reviewed October 8, 2026. Review dates are maintained as described on the About page.