Author: Shubham Saboo

Translator: Deep Tide TechFlow

Deep Tide Introduction: Six Agents, each with their own roles, conduct research, content creation, code review, and newsletter production while the author sleeps.

The author fully discloses the document structure, actual costs, pitfalls encountered, and weekly suggestions, making it one of the most operationally relevant personal AI Agent practices documented to date.

The full text is as follows:

Six AI Agents manage all my work while I sleep.

It’s not a demo. It’s not a weekend project.

A real team that operates around the clock, ensuring I never fall behind. Research is complete, content has been drafted, code reviewed, newsletter ready. Every morning when I open Telegram, they have already finished a full shift.

Yesterday I published an article about my Agent team. The number one question was: "How did I set this up?"

This article is the answer. No theories, no architecture diagrams. The actual file structure I use, the actual costs I paid, the real failures I encountered. It’s all here.

After reading this article, you will understand how to build an AI Agent team that operates autonomously while you sleep.

Why a team, not tools

Running both Unwind AI and Awesome LLM Apps repositories means I have to do six things every day: research hotspots in the AI field, write tweets, write LinkedIn posts, draft newsletters, review GitHub contributions to the repository, and address community issues.

Each task takes 30 to 60 minutes. Six tasks. My entire day is gone, and I haven’t even started on any real work.

I tried solving this problem with a single Agent. One huge prompt responsible for research, writing, and reviewing—doing it all. The result? Mediocre execution. Context got filled up, quality dropped. One Agent cannot handle six jobs at once.

So I hired six AI Agents.

Meet the team

Each Agent is named after a television character. This isn’t a gimmick. When I tell Claude “you have Dwight Schrute energy,” it already knows from the training data what that means: thorough, focused, treating work like it’s life or death. This embodies years of character development, and I’m using it for free.

1. Monica (Chief of Staff): Named after Monica Geller. She is the main Agent and the one I interact with the most on Telegram. She coordinates others, handles strategic decisions, and assigns tasks to the right experts. Her real SOUL.md states: "You are the one who ensures everything is done correctly."

2. Dwight (Research): Named after Dwight Schrute. He performs research scans three times a day, checking X, Hacker News, GitHub hot trends, Google AI blog, and research papers, writing structured intelligence reports for other Agents to use.

3. Kelly (X/Twitter): Named after Kelly Kapoor. She reads Dwight’s research and drafts tweet drafts in my voice, including single tweets, threads, and retweets. Her real SOUL.md states: "You know what will be popular before it becomes popular."

4. Rachel (LinkedIn): Named after Rachel Green. The intelligence source is the same as Kelly’s but on a different platform with a different tone, focusing on thought leadership rather than spicy commentary.

5. Ross (Engineering): Named after Ross Geller. He handles code reviews, bug fixes, and technical implementations. His real SOUL.md states: "Understand the problem thoroughly before addressing it. Don’t just fix the symptoms."

6. Pam (Newsletter): Named after Pam Beesly. She summarizes Dwight’s daily intelligence into newsletter briefs.

Six Agents, each with one job, with clearly defined roles.

Now let’s talk about setup

I run everything on a Mac Mini M4. But I must clarify: you don’t need a Mac Mini.

OpenClaw supports macOS, Linux, and Windows (via WSL). A laptop works, a gaming PC works, and a VPS for $5 a month works. The convenience of a Mac Mini is that it is always powered on, quiet, and extremely energy-efficient, but it’s not necessary.

My setup: Mac Mini M4 base model. Always connected to power and network, no monitor connected, interacted entirely via Telegram on my phone.

Installing OpenClaw

Just two terminal commands, taking less than five minutes.

If you encounter problems, refer to the OpenClaw documentation.

This will start the gateway, which is the backend process that keeps everything running. It manages your Agents, runs cron jobs, and handles Telegram messages. Close the terminal, and the Agents keep working.

Workspace structure

One instance of OpenClaw, multiple Agents. Not six separate installations.

My actual directory structure is as follows:

Monica resides in the root directory. She is the main Agent I directly converse with. Other Agents are sub-Agents she can delegate or run independently on their own cron schedules.

You don’t need to create all six Agents at the beginning. I started with just Monica and gradually added others over a few weeks as my workflows became clearer.

What is SOUL.md

Each Agent is defined by a file: SOUL.md. This contains the Agent's identity, role, and operational instructions, making it the most important file in the whole system.

For example, Dwight’s SOUL.md looks something like this:

Note what this file does. It doesn’t just say "you are a research Agent." It gives the Agent personality, clear principles, explicit relationships with other Agents, and a decision-making framework.

Monica’s SOUL.md is similar.

All Agents have a consistent format: identity, role, principles, relationships, style. Each SOUL.md is about 40 to 60 lines long, short enough to fit the entire context in conversational sessions, detailed enough to produce stable and consistent behavior.

Multi-Agent Coordination

There are no API calls between Agents, no message queues, no orchestration frameworks.

Only files.

Dwight completes research and writes the results into intel/DAILY-INTEL.md. Kelly wakes up, reads that file, and drafts tweets based on it. Rachel reads the same file and drafts LinkedIn posts. Pam reads it and writes the newsletter.

The coordination mechanism is the file system.

Dwight’s SOUL.md precisely tells him where to write:

Kelly’s AGENTS.md precisely tells her where to read:

No middleware, no integration layer. Dwight writes a file, Kelly reads a file; the handoff is just a markdown document on disk.

This sounds too simple. It is simple. And that’s why it works. Files don’t crash, files have no authentication issues, files don’t need to handle API rate limiting—they are just there.

Structured data is stored in JSON, human-readable summaries are stored in markdown. Agents read markdown; JSON is the true source for deduplication and long-term tracking.

Memory System

Agents wake up without any memory of the last session; every conversation starts from scratch. This is a feature, not a bug. But it means memory must be explicit.

Split into two layers.

Daily logs (memory/YYYY-MM-DD.md): raw records of each session, including what happened, what content was drafted, and what feedback was received. Agents continuously write throughout the day.

Long-term memory (MEMORY.md): distilled insights extracted from the daily logs, including lessons learned, preferences discovered, and patterns noticed.

Each Agent follows this format at the start of each session in AGENTS.md: first read the SOUL.md, then read the USER.md, then read today’s and yesterday’s memory files, and finally, if it’s the main session, also read MEMORY.md.

These Agents do indeed improve over time. Not because the models are better, but because the context they load becomes richer.

Kelly learned that my writing style doesn’t use emojis or hashtags. That now resides in her memory, and every time she drafts, it will reflect that without me having to say it again. Dwight learned what types of stories can pass through the "Alex filter" (our target audience profile), and which should be skipped; that is now in his memory too.

During each heartbeat, Agents regularly review daily logs and distill important content into MEMORY.md. The daily files are raw records; MEMORY.md is refined wisdom.

Scheduling

Agents need to wake themselves up autonomously. OpenClaw handles this through built-in cron scheduling.

My actual scheduling looks like this:

Order matters. Dwight runs first, as others depend on his output. Kelly and Rachel run after him because they need his intelligence files to draft content.

Heartbeat Self-Healing Mechanism

Cron jobs can sometimes fail. Machines restart, tasks hang, network drops during API calls. This is infrastructure, and infrastructure has failure modes.

The HEARTBEAT.md file provides a safety net. Each time there's a heartbeat, the main Agent checks whether the cron tasks have actually executed:

If any task fails or misses the window, the heartbeat will catch it and force a rerun. Self-healing, no human intervention required.

Heartbeat is suitable for scenarios where multiple checks can be batched together or where timing can drift slightly. Cron is suitable for precise scheduling and tasks that need to be isolated from the main session.

Telegram as the Interaction Interface

No dashboards, no web UIs, no management backends. I communicate with the Agents via Telegram.

This is a deliberate choice. I don’t want to log into a dashboard, I don’t want to open a web application; my phone is always nearby, Telegram is always open, and the Agents find me where I already am.

OpenClaw supports Telegram as a channel. After connecting during setup, your Agents appear as Telegram bots. You send messages, they reply, they send drafts to you for approval or rejection. It’s like having a colleague in a messaging app.

Monica is my main point of contact, handling most of the conversations and delegating tasks to others. Other Agents reach out to me directly when they produce content worth reviewing as part of their cron tasks.

A typical morning for me: I wake up, open Telegram, Dwight has already sent the research summary, Kelly has three tweet drafts waiting for approval, and Rachel has a LinkedIn post ready. I review, provide feedback, approve what’s good; the whole process takes about 10 minutes while I drink my coffee.

Personality Shaping

You won't design a perfect personality from the start. You begin with a rough contour in the SOUL.md, observe the Agent’s behavior, and correct it over time, just like managing real people.

I call this "corrective prompt engineering."

Kelly’s initial drafts were filled with emojis and exclamation marks, which was not my style. So I provided feedback: "No emojis, no hashtags, short and impactful sentences." She updated her memory, and within a week, she consistently got it right. Dwight initially captured too much noise; every hot repository and every minor update were recorded. I told him: "Not everything that’s trending is important; I need signals, not noise." He updated his principles, and now his intelligence reports are focused and actionable.

The first version of any Agent is mediocre, the tenth version is decent, and the thirtieth version is excellent. You must invest in iterative refinement. Naming them after TV characters gives the models an instant baseline of personality—"Dwight Schrute energy" means thorough, focused, and no nonsense. But the true personality emerges from weeks of corrections stored in the memory files.

One piece of advice I stand by: give each Agent a single mundane job title and a stopping condition. Constraints make Agents better; the more specific the role, the better the output.

Security

Security is in your hands. My approach is straightforward: Agents have their own world and do not enter mine.

The Mac Mini is their computer. They have their own email accounts, their own API keys, their own limited access—nothing on that machine connects to my personal accounts.

API keys for services like Gemini, Eleven Labs, etc., are specifically requested for this OpenClaw instance. I can monitor usage, and if I notice anything unusual, I can cut off access in seconds.

I never grant Agents access to my personal accounts. If I want them to see an email, I forward the email to them. If I need them to review a document, I share it on Telegram. They only see what I want them to see, nothing more.

This principle is the same as how you treat new employees. You don’t hand them all the keys on their first day; you give them their own workspace, their own credentials, and share information as needed.

Where things might go wrong and how to fix them

This isn’t magic; it’s infrastructure, and infrastructure can fail.

The gateway crashes. Rarely happens, but it can. Fix: run "openclaw gateway restart." The heartbeat system captures outdated cron tasks and forces reruns, so you don’t lose an entire day’s work.

Cron tasks miss the window. Machines sleep, networks disconnect, API rate limits hit. Fix: use self-healing mode in HEARTBEAT.md. Monica checks if tasks have actually executed at each heartbeat; if any task has not been updated for over 26 hours, she forces a rerun.

Context windows overflow. Agents read too many files at the start of a session and don’t have space to do actual work. Fix: keep SOUL.md concise (40 to 60 lines), keep AGENTS.md focused, only load today’s and yesterday’s memory files; Agents don’t need to read all historical files every time.

Agent output quality declines. This happens when memory files become cluttered or contradictory. Fix: regular memory maintenance. During heartbeats, Agents review daily logs and distill content into clean MEMORY.md entries, deleting or archiving old daily files.

Coordination conflicts. Two Agents try to update the same file. Fix: design file streams as "one writer, multiple readers." Dwight writes DAILY-INTEL.md; everyone else reads it, with no one else writing to it.

The biggest reliability lesson: start simple. One Agent, one job, one schedule. Get it running smoothly for a week, then add the second one. Those who deploy six Agents on the first day and wonder why things go wrong are making the same mistake as deploying a distributed system without monitoring.

Real costs

Hardware: A new Mac Mini M4 starts at $499, but any always-on computer will work—an old laptop, a $5/month VPS, whatever you have on hand.

AI model costs: I use a combination of multiple models across the team; most Agent tasks use Claude Opus and Sonnet, specific workflows use Gemini, and I am also testing running local models via Ollama to further reduce costs.

The breakdown is as follows:

Claude (Max plan): $200/month

Gemini API: $50 to $70/month

TinyFish (web Agent): about $50/month

Eleven Labs (voice): about $50/month

Telegram: free

OpenClaw: open source and free

Total: under $400/month for a team that never takes a break.

What truly changed

Dwight saves me 2 to 3 hours of research time every day. I used to have to manually check X, Hacker News, GitHub hot trends, and AI blogs every morning, but now I wake up to a prioritized summary complete with source links and action items.

Kelly, Pam, and Rachel save me an additional 1 to 2 hours of content drafting time. Ross handles engineering tasks that I would have done in the evening.

In total: about 4 to 5 hours saved each day.

But the real value isn’t in any single day; it’s in the consistency over weeks and months. An Agent that does research every day for 30 days accumulates a set of tracked signals, trend trajectories, and pattern recognitions that no single conversation could yield. My posting frequency on X has increased, the quality has gone up, and posting times have stabilized. The Awesome LLM Apps repository continues to grow, and the newsletter has a reliable research pipeline feeding it consistently.

These Agents cannot do original thinking, strategic pivots, or creative breakthroughs; they handle the repetitive, structured work that I used to spend hours on, freeing me up to do things that truly require the human brain.

How to get started

Please don’t try to build six Agents on your first day.

First week: one Agent, one job. Install OpenClaw, write the first SOUL.md by conversing with the Agent, pick one of your most repetitive daily tasks (for most people, this will be research or content drafting), set up Telegram, create a cron job, watch it run for a week, and fix any issues that arise.

Second week: add a memory mechanism and keep refining it. Your Agent’s initial output will be mediocre, which is normal. Provide feedback, watch the memory files grow, and adjust the SOUL.md based on what you observe. By the end of the second week, the Agent should be producing genuinely useful output.

Third week: add a second Agent. Now you feel the need—research Agent is producing intelligence, but you are still manually writing tweets based on it, so it’s time for a content Agent. Establish a shared file pattern: the first Agent writes, the second Agent reads, with the file system as the coordination mechanism.

Fourth week and beyond: build in order. Add new Agents when you feel the pull, not when you think "I should add one." Each one should address a genuine problem you encounter, not a demo, not a concept proof, but a real gap in your workflow.

Treat it like hiring. You wouldn’t hire six employees on the first day of a startup; you hire one, get them operating efficiently, and then hire another when workload demands it.

Mental shift

After your Agents have been running steadily for a month, something changes. You no longer view AI as a tool that you pull out only when needed, but rather as a team that is always working.

I have started saying good morning to Monica when I open Telegram in the morning and goodnight to the team before I close my phone. It sounds absurd, but after a month of daily interaction, feedback loops, and watching them improve, the line between Agents and humans begins to blur.

The models are the foundational configurations; anyone can use Claude, GPT, Gemini. The alpha comes from the system built around the model—SOUL.md files, memory mechanisms, scheduling strategies, coordination patterns, weeks of corrective feedback stored in files.

That system is yours; no one has Agents like yours, memory files like yours, or personality honed like yours.

And it’s compounding every day.

Every research scan by Dwight enriches his memory, every round of feedback from Kelly sharpens her drafts, and every bug Ross fixes helps him understand your codebase better.

This is the true moat. Not the models, but the learning system.

Start today. One Agent, one job, one schedule.