An AI Agent for Productivity That Actually Runs Your Day

Most people use AI like a vending machine. Ask a question, get an answer, close the tab. That is a chatbot. It is not leverage.

The thing that changes your day is an agent that already knows your stack, holds your rules in memory, runs jobs while you sleep, and hands you finished work instead of suggestions. You stop being the person who writes every prompt and become the person who approves the output. Set up right, a private AI agent for productivity reclaims the first two hours of your morning and the last hour of your night. This is how to build that, using the exact patterns operators are running on Hermes today.

Start with delegation, not conversation

Here is the mental shift that matters. Before you ask the agent to do anything, ask yourself: is this something only I can do, or something I am doing out of habit?

Alex Finn put it well after running this setup for months: "Think about what the agent can do versus what only you can do. The board is like having an employee. You start your day, give it all the tasks, go get your human work done."

In practice that looks like a morning triage ritual. You transcribe your to-do list into a shared board, the agent claims everything it can handle, and you keep the items that genuinely need a human. An hour later you come back to finished work. The board is the handoff point, not a chat thread, so nothing gets lost when a session ends. If you want to run several specialists at once instead of one generalist, that is its own discipline, covered in our guide to multi-agent orchestration.

Let the agent train itself on you

A generic assistant stays generic. A good one compounds. The trick is a tiny daily ritual that feeds the agent a fresh signal about what you actually care about.

Set a cron for 9am that asks one question: "What is your number one priority today?" Your answer does two things. It tells the agent your current focus so it can suggest tasks that move that priority, and it updates the agent's long-term memory about you. Run it for three weeks and the agent stops guessing. It starts pre-loading work that matches where your head is.

Every morning at 9am, ask me what my number one priority is today.
Then list tasks you can take off my plate to move that priority forward,
and update your memory of my current focus accordingly.

The reason this works is that your real intentions live in what you say day to day, not in a profile doc you wrote once and forgot. Memory is the engine here, and it is worth setting up properly. See how agent memory is structured for the tiers that keep this from bloating your context.

The dreaming brief: do the thinking at 6am, deliver it at 7am

The best morning brief pattern splits reasoning from delivery, because they cost different amounts.

Run two crons. The 6am job uses a cheap model to read everything: recent conversations, meeting notes, your coding logs, whatever context it has. It synthesizes three recommendations and one non-negotiable for the day and writes them to a file. The 7am job reads that file, adds today's weather and calendar, and sends you the brief.

Why bother with two jobs? Because heavy reasoning is expensive and you do not want it happening while you are standing there waiting. Jack Roberts, who runs this daily, described it plainly: "The dreaming sequence at 6am synthesizes heavy reasoning into a file. The 7am brief just reads that file. Fast, cheap, and always fresh." Assign the overnight job a cheap model and the savings stack up over a month. The pre-computed file is also readable by any later job, so the reasoning gets reused instead of repeated. There is more on scheduling these jobs cleanly in our cron automation guide.

Become the approver, not the author

Email is the obvious win, and it is also where people make the obvious mistake. They let the agent send.

Do not. Set up triage as draft-only. One cron runs twice a day and drafts replies to your unread mail. A second cron at 8am delivers a prioritized brief: your high-priority emails with deep links straight to the thread, your calendar, and anything else you want on the dashboard. You read, you tweak, you hit send. The agent never touches the outbound button.

This is not a limitation, it is the whole point. As Jack Roberts said about his triage setup: "Draft is what I like, then I manually approve it. You cannot have a single percent of a failure rate." A wrong calendar summary costs you nothing. A wrong email sent to a client costs you the client. Keep the human gate where the stakes are real. The same draft-only discipline underpins safe agent design generally, which we cover in AI agent security.

One more thing on crons: token-budget them. Do not run triage every ten minutes. Twice a day is plenty, and you will feel the difference on your bill.

Use the right connection, in the right order

When you wire the agent into your tools, there is a hierarchy that saves you hours of flaky behavior. Jack Roberts calls it the connection priority matrix:

• First choice: an MCP or API connector. Fastest and most reliable. Always check whether the tool has one before doing anything else.
• Second choice: a local file the agent reads and edits directly.
• Last resort: screen control, where the agent drives the cursor by looking at the screen.

Most people reach for screen control first because it feels intuitive. Invert that. "If you can connect it via an MCP, do that. It will be a lot faster than using the computer. Only as a last resort do we want desktop intelligence." Screen control is slow and breaks when a button moves. An API does not.

A few non-obvious moves worth stealing

A handful of smaller techniques punch above their weight:

• Meta-prompt your long builds. Instead of writing a /goal prompt yourself, ask the agent to write the perfect /goal prompt first, then run its output. The detail and guardrails it produces beat what you would write by hand, and it is the difference between a goal that runs for 24 hours productively and one that wanders.
• Run live competitor breakdowns. "Open a browser, go to [competitor URL], do a full technical breakdown. Stack, features, analytics events, third-party scripts. Save as markdown." Then feed that file straight into your build agent.
• Self-terminating monitoring. "For the next 12 hours, check this page every 10 minutes and respond. Kill the cron when the time is up." You manage the whole lifecycle in one sentence.
• Pair your devices. With a private network across your machines, the agent can check whether localhost:3000 is running the latest build on your other computer, pull files, or run a local model on one machine and use the result on another.

FAQ

What is an AI agent for productivity, exactly?

It is a persistent assistant that works through your real tools, holds your preferences in memory, and runs scheduled or triggered jobs on its own. Unlike a chatbot, it keeps state between sessions and produces finished work you approve rather than answers you copy and paste.

How is this different from ChatGPT or a normal assistant app?

A chatbot forgets you the moment you close the tab and cannot act on a schedule. An agent remembers your rules, drafts and triages on a cron, connects to your email and calendar, and improves as you correct it. You delegate to it instead of querying it.

Will it send emails or take actions without my approval?

Only if you let it, and the recommended setup does not. Keep high-stakes actions like sending email draft-only and put a human approval gate in front of anything irreversible. Low-stakes work like briefs and research can run fully autonomously.

How long before it actually feels personalized?

About two to three weeks of a daily priority prompt plus normal use. Memory compounds, so the agent that felt generic in week one starts pre-loading the right work by week three.

Where to start

You do not need to build all of this at once. Pick the morning brief, wire the daily priority prompt, and set email triage to draft-only. That alone changes how your day starts.

If you would rather skip the trial and error, OpenClawCrew sells private AI agent starter kits for $49 that ship with these patterns already wired, and a done-for-you setup service for teams that want the whole thing standing the first week. Either way you end up with an agent that runs on Hermes (or OpenClaw), remembers your rules, and works through your tools. Once your personal ops are humming, the natural next step is to point the same agent at shipping real software. See how that works in building AI agent apps.