Set Up Your Personal Agent
This content was developed with AI assistance and is regularly reviewed for accuracy.
Snapshot: October 2026. Model names, prices, and product features change monthly; What's New in AI is the maintained chronology.
A personal agent is an assistant that knows your work, can reach the tools you use, and does recurring jobs before you ask. A couple of years ago, building one meant writing a folder of configuration files and a script to wake it up. In 2026 every piece of that setup has a home inside the assistant you already use. This playbook maps the old pieces onto projects, memory, skills, connectors, and scheduled tasks, then sets boundaries so the agent works ahead of you without acting beyond what you approved.
Last tested with Claude (Pro plan: projects, memory, skills, connectors, scheduled tasks) and ChatGPT (Plus plan: projects, memory, scheduled tasks) in October 2026.
What you'll accomplish
- Choose three recurring jobs and set clear limits on what the agent may do
- Give it identity and context through project instructions and memory
- Connect only the tools those jobs require
- Add skills for the procedures it repeats
- Run a daily heartbeat as a scheduled task, and expand one job at a time
Prerequisites
- The earlier intermediate playbooks, or equivalent: a brief and skill library and at least one scheduled report
- Claude Pro or ChatGPT Plus (or above). Free plans cover Steps 1-4; the heartbeat in Step 5 needs scheduled tasks.
- 60-90 minutes for setup, then 10 minutes a week
Other routes exist. Google's Gemini Spark (announced May 2026) runs as an always-on agent in Google's cloud for US AI Ultra subscribers. ChatGPT's Dots, persistent agents announced in September 2026, are reported to be on Pro and Business Premium tiers. Meta launched a Muse personal-agent app in the US in September 2026. OpenClaw is an open-source agent you run on your own machine with your own model keys; it is powerful and needs broad permissions, so treat it as a power-user project. The principles below apply to all of them.
From config files to built-in features
If you have seen guides that build an "agent operating system" from files, here is where each piece lives now:
| Old piece | What it held | 2026 home |
|---|---|---|
| IDENTITY.md | Role, tone, boundaries | Project instructions |
| USER.md | Who you are, how you work | Memory and personal preferences |
| TASKS.md | The current task list | A doc the agent reads through a connector |
| skills/ folder | Procedures for specific jobs | Agent Skills (SKILL.md) |
| TOOLS.md | What the agent can use | Connectors |
| Heartbeat script | Wake the agent on a timer | Scheduled tasks |
The built-in versions are better in one important way: they carry permissions. A connector shows you what it can read and write; a scheduled task lists when it runs. That makes the boundaries in Step 1 enforceable instead of aspirational.
The playbook
Step 1: Choose three jobs and set the limits
Goal: Decide what the agent does, and what it never does, before you give it access.
Action: Pick three recurring jobs. Good first choices read and draft: a morning briefing, meeting prep, the weekly report. Then write the limits down, because they shape every later step:
May read: my calendar, my email, the "Tasks" doc, my project files.
May draft: emails (as drafts), documents, plans, summaries.
Must ask first: creating calendar events, moving or renaming files.
Never: send email or messages, delete anything, accept invitations, buy anything, share documents.
The useful question is "what could this touch if it got something wrong?" The answer is its blast radius. Agents that only read and draft have a small one. The data-loss incidents of 2026 mostly involved agents that held far more access than their task needed.
Checkpoint: Three jobs and a written list of what the agent may read, may draft, must ask about, and must never do.
Step 2: Give it identity and context
Goal: An agent that already knows your role, standards, and priorities.
Action: Create a project called "Personal agent" and put identity and preferences in its instructions:
You are my chief of staff. I'm a [role] at [organization], in [time zone].
Priorities this quarter: [two or three].
Style: lead with the decision or ask; bullet points; under 200 words unless I ask for more.
Deep work is 8-11am; protect it.
Follow the limits below without exception: [paste your Step 1 list].
When unsure, ask one clarifying question rather than guessing.
Then review memory. Claude includes memory on every plan and, reportedly, does not store sensitive topics unless you opt in. ChatGPT's memory is off by default in the EU, UK, and Switzerland. Open the memory settings, read what is stored, delete anything wrong or anything you would not want used, and add the facts it should always know. Instructions are what you tell it on purpose; memory is what it picks up. Check both.
Checkpoint: In a new conversation inside the project, ask "What do you know about how I work?" The answer is accurate and contains nothing you would not want there.
Step 3: Connect only what the jobs need
Goal: Give the agent its tools without widening the blast radius.
Action: Go through your three jobs and connect only the services they read. For most people that is calendar, email, and a document store. Put your task list in a single doc (call it "Tasks") that the agent reads through the Drive connector, with sections for this week, waiting on others, and recurring items.
When you approve each connector, read the permissions. If it can send or delete, your Step 1 limits are now the only thing stopping it, so keep them in the project instructions. For anything sensitive, test with low-stakes data first. Anthropic's own guidance for connectors is to verify every one before you connect it, and to remove the ones a job does not use.
Checkpoint: The agent can read your calendar, email, and Tasks doc, and nothing else.
Step 4: Add skills for repeated procedures
Goal: Consistent output for the jobs the agent repeats.
Action: Add a skill for each job's procedure, using the SKILL.md format from Build Your Brief and Skill Library. A morning-briefing skill might say: list today's meetings with one line of prep each; flag anything due today or overdue in Tasks; surface emails that need a reply from me today; end with a suggested top three. Keep each skill short and give it a description that names when to use it.
Checkpoint: Ask "brief me for today" in the project. The output follows your skill without you describing the format.
Step 5: Run a daily heartbeat
Goal: The agent does its first job before you open your laptop.
Action: A heartbeat is simply a scheduled task. Set one in the project:
Every weekday at 7am: Run my morning briefing. Read today's calendar, my Tasks doc, and email received since yesterday at 5pm. Follow the morning-briefing skill. Save the briefing as a draft email to myself. Do not send, reply to, archive, or delete anything.
Add a weekly one on Monday to review the Tasks doc: what carried over twice, what is blocked, what to drop.
Remember that the heartbeat reads text other people wrote. An email or invite can contain instructions aimed at your agent, and 2026 security research showed agents hijacked exactly this way. The read-and-draft limits are your protection: an agent that cannot send or share has nothing useful to be tricked into doing.
Checkpoint: Three mornings of briefings arrived, each accurate against your calendar and inbox.
Step 6: Review weekly and expand slowly
Goal: Grow what the agent does without losing track of it.
Action: Spend ten minutes each week on three questions. Was each output right? Is any scheduled task no longer read? Do the instructions, memory, or Tasks doc need updating? Add one new job only after the current ones have run cleanly for two weeks. Each scheduled run uses your plan's limits, so prune what you have stopped reading.
When you are ready to let the agent take actions rather than just draft, raise one permission at a time, keep approvals on, and read Agent Safety and Guardrails first.
Checkpoint: After a month, three or more jobs run reliably, and you can list every scheduled task and connector the agent has.
Common pitfalls
- Granting access before setting limits: Write the never-do list first; it decides which connectors you approve.
- Never reading memory: Stale or wrong memories quietly skew every answer. Review them monthly.
- Adding five jobs in week one: You cannot check five new outputs well. Add one at a time.
- Letting it send on your behalf too early: Drafts are reversible. Sent messages, deletions, and purchases are not.
- Forgetting what is scheduled: Keep a list of every scheduled task and delete the ones you ignore.
Key takeaways
- Every piece of an old agent config now has a built-in home: instructions, memory, connectors, skills, scheduled tasks
- Set what the agent may read, draft, ask about, and never do before you connect anything
- Connect only the tools your chosen jobs need, and read each connector's permissions
- A heartbeat is a scheduled task; keep it read-and-draft
- Review weekly, prune what you ignore, and expand one job at a time
Next steps
Continue to: Agents in the Apps You Already Pay For for what each assistant's agent mode can do and how its approvals work.
Continue to: Loops and Harnesses for agents that run toward a goal and check their own work.
Continue to: Protect: Privacy, Permissions, Injection, Irreversible Actions for the full set of safeguards before you widen permissions.