Google Senior AI Product Manager: 6 agents handle my daily tasks, costing less than $400 per month, running 24/7

After reading this, you’ll understand how to build an autonomous AI agent team that operates while you sleep.

Author: Shubham Saboo

Translation: Deep潮 TechFlow

Deep潮 Guide: Six specialized agents complete research, content creation, code review, and newsletter production while the author sleeps.

The author fully discloses the file structure, actual costs, pitfalls encountered, and weekly recommendations, making this one of the most practically valuable personal AI agent implementation records to date.

Full Text:

Six AI agents manage all my work while I sleep.

This is not a demo. Not a weekend project.

A truly 24/7 operational team ensures I never fall behind. Research is done, drafts are ready, code reviews are complete, newsletters are prepared. By the time I open Telegram in the morning, they’ve already completed a full shift.

Yesterday, I published an article about my agent team. The top question was: “How exactly did I build this system?”

This is the answer. No theory, no architecture diagrams. I’m sharing the actual file structure I use, the real costs I’ve paid, the failures I’ve encountered. Everything is here.

After reading this, you’ll understand how to build an autonomous AI agent team that runs while you sleep.

Why a team, not tools

Running Unwind AI and the Awesome LLM Apps repositories simultaneously means doing six things every day: tracking AI industry trends, tweeting, posting on LinkedIn, drafting newsletters, reviewing GitHub contributions, and handling community issues.

Each task takes 30 to 60 minutes. Six tasks. That’s my entire day gone before I even start real work.

I tried solving this with a single agent. A massive prompt handling research, writing, and review—everything bundled. The result? Mediocrity. The context filled up, quality dropped. One agent can’t handle six jobs at once.

So I hired six AI agents.

Meet the team

Each agent is named after a TV character. This isn’t a gimmick. When I tell Claude, “You have Dwight Schrute’s energy,” it already knows what that means from training data: thorough, focused, treating work as life. This comes from 30 seasons of character traits I’ve leveraged for free.

1. Monica (Chief of Staff): Named after Monica Geller. She’s the main agent, the one I interact with most on Telegram. She coordinates others, handles strategic decisions, assigns tasks to specialists. Her SOUL.md states: “You are the one ensuring everything gets done right.”

2. Dwight (Research): Named after Dwight Schrute. Performs research scans three times daily, checking X, Hacker News, GitHub trending, Google AI blogs, and papers, then writes structured intelligence reports for all agents.

3. Kelly (X/Twitter): Named after Kelly Kapoor. Reads Dwight’s research, drafts tweets in my tone, including single tweets, threads, and quote tweets. Her SOUL.md states: “You knew it would trend before it was trending.”

4. Rachel (LinkedIn): Named after Rachel Green. Uses the same intelligence sources as Kelly but with a different platform and tone—focused on thought leadership rather than hot takes.

5. Ross (Engineering): Named after Ross Geller. Handles code reviews, bug fixes, and technical implementations. His SOUL.md states: “When solving problems, understand thoroughly first. Don’t just fix symptoms.”

6. Pam (Newsletter): Named after Pam Beesly. Organizes Dwight’s daily intelligence into newsletter summaries.

Each agent has a dedicated role, with clear division of labor.

Now, how to build it

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). Laptops work, gaming PCs work, $5/month VPS works. The Mac Mini’s advantage is always-on power, quiet operation, and low energy use—but it’s not mandatory.

My setup: Basic Mac Mini M4. Always plugged in, connected to the internet, no display needed, interacting entirely via Telegram on my phone.

Installing OpenClaw

Just two terminal commands, less than five minutes.

If you encounter issues, consult the OpenClaw documentation.

This launches the gateway, the background process that keeps everything running. It manages your agents, runs cron jobs, handles Telegram messages. Close the terminal, and the agents keep working.

Workspace structure

One OpenClaw instance, multiple agents. Not six separate installs.

My actual directory structure looks like this:

Monica lives in the root directory. She’s the main agent I interact with directly. Other agents are delegated sub-agents or run independently on their own cron schedules.

You don’t need to set up six agents from the start. I began with just Monica, and as my workflow became clearer, I gradually added others over several weeks.

What is SOUL.md?

Each agent is defined by a file: SOUL.md. This is the agent’s identity, role, and operational instructions— the most critical file in the system.

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

Note what this file does. It’s not just saying “You are a research agent.” It endows the agent with personality, clear principles, explicit relationships with other agents, and a decision-making framework.

Monica’s SOUL.md is similar.

All agents follow the same pattern: identity, role, principles, relationships, style. Each SOUL.md is about 40 to 60 lines—short enough to fit into context per session, detailed enough to produce consistent, stable behavior.

Coordination among multiple agents

Agents don’t call APIs to each other, don’t use message queues, don’t have orchestration frameworks.

They only use files.

Dwight completes research and writes results into intel/DAILY-INTEL.md. Kelly wakes up, reads that file, and drafts tweets accordingly. Rachel reads the same file to draft LinkedIn posts. Pam reads it to compile newsletters.

Coordination is managed via the filesystem.

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

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

No middleware, no integration layer. Dwight writes a file, Kelly reads a file—handover is a markdown document on disk.

It sounds too simple. It is. That’s why it works. Files don’t crash, don’t have authentication issues, don’t need API rate limiting. They’re just there.

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

Memory system

Agents have no memory of previous conversations—they start fresh each time. This is a feature, not a flaw. But it means memory must be explicit.

Two layers:

Daily logs (memory/YYYY-MM-DD.md): raw records of each session, including what happened, drafts created, feedback received. Agents continuously write into these throughout the day.

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

At session start, each agent follows instructions in AGENTS.md: read SOUL.md, then USER.md, then today’s and yesterday’s memory files, and if it’s the main session, also MEMORY.md.

These agents do improve over time—not because the models change, but because their loaded context becomes richer.

Kelly learns my writing style, no emojis or hashtags needed. That’s stored in her memory, so every draft reflects it without me reminding her. Dwight learns which stories pass the “Alex filter” (our target audience profile) and which to skip—also stored in his memory.

During each heartbeat, agents periodically review daily logs, extract key content into MEMORY.md. Daily files are raw records; MEMORY.md is distilled wisdom.

Scheduling

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

My actual schedule:

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

Heartbeat self-healing

Cron jobs sometimes fail—machine reboots, tasks hang, network issues during API calls. Infrastructure can fail.

HEARTBEAT.md provides a safety net. Each heartbeat, the main agent verifies whether cron tasks actually executed:

If a task failed or missed its window, the heartbeat detects it and forces a rerun. Self-healing without manual intervention.

Heartbeat is suitable for batch checks or scenarios with slight timing flexibility. Cron is better for precise scheduling and tasks isolated from main sessions.

Telegram as interface

No dashboards, no web UI, no management backend. I interact with agents via Telegram.

This is deliberate. I don’t want to log into dashboards or web apps. My phone is always nearby, Telegram is always open, and agents find me where I am.

OpenClaw supports Telegram as a channel. After setup, your agent appears as a Telegram bot. You send messages, it replies, drafts are sent for approval or rejection—like having a colleague in your chat app.

Monica is my primary contact, handling most conversations and delegating tasks. Other agents contact me directly when their cron outputs are ready for review.

My typical morning: wake up, open Telegram, Dwight has already sent a research summary, Kelly has three tweet drafts awaiting approval, Rachel has a LinkedIn post ready. I review, give feedback, approve—all in about 10 minutes with coffee.

Personality shaping

You don’t start with a perfect personality. Begin with a rough outline in SOUL.md, observe behavior, and refine over time—like managing real people.

I call this “corrective prompt engineering.”

Kelly’s initial drafts are full of emojis and exclamation points—that’s not my style. I give feedback: “No emojis, no hashtags, short and punchy sentences.” She updates her memory, and after a week, she’s doing it right. Dwight initially captured too much noise—every trending repo, every small update. I told him, “Not everything trending is important. I need signals, not noise.” He revised his principles, and now his reports are focused and actionable.

Any agent’s first version is mediocre; the tenth is decent; the thirtieth is excellent. You must iterate and refine. Naming them after TV characters gives the model an instant personality baseline—“Dwight Schrute energy” means thorough, focused, no nonsense. But true personality emerges from weeks of stored corrections in memory.

A good tip: give each agent a simple, mundane job title and a clear stop condition. Constraints improve performance; more specific roles yield better output.

Security considerations

Security is in your hands. My approach is simple: agents have their own environment, separate from mine.

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

API keys for services like Gemini, Eleven Labs, etc., are dedicated to this OpenClaw instance. I monitor usage, and if anomalies appear, I can cut access instantly.

I never give agents access to my personal accounts. If they need to see an email, I forward it. For documents, I share via Telegram. They only see what I want them to see.

Treat it like onboarding new employees: don’t give them all keys on day one. Provide their own workspace, credentials, and share information as needed.

Where issues can arise and how to fix

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

Gateway crashes. Rare but possible. Fix: run “openclaw gateway restart.” The heartbeat system detects outdated cron tasks and forces reruns, so no work is lost.

Cron misses a window. Machine sleeps, network disconnects, API rate limits. Fix: HEARTBEAT.md’s self-healing mode. During each heartbeat, Monica checks if tasks actually ran; if any haven’t updated in over 26 hours, she forces a rerun.

Context window overflow. Agents read too many files at session start, leaving no room for actual work. Fix: keep SOUL.md concise (40-60 lines), focus AGENTS.md, load only today’s and yesterday’s memory files. Agents don’t need to read all history every time.

Output quality drops. Happens when memory files become cluttered or contradictory. Fix: regular memory maintenance. During heartbeat, agents review daily logs, distill key insights into MEMORY.md, and archive or delete old daily files.

Coordination conflicts. Two agents try to update the same file simultaneously. Fix: design file flow as “one writer, multiple readers.” Dwight writes DAILY-INTEL.md; others read it. No one writes to it except Dwight.

The biggest lesson for reliability: start simple. One agent, one task, one schedule. Make it stable for a week, then add a second. Those who deploy six agents on day one and wonder why it breaks are making the same mistake as deploying a distributed system without monitoring.

Actual costs

Hardware: Mac Mini M4 starts at $499, but any always-on computer works—old laptops, $5/month VPS, whatever you have.

AI model costs: I use multiple models across the team. Most tasks use Claude Opus and Sonnet; some workflows use Gemini. I’m also testing local models via Ollama to cut costs further.

Breakdown:

Claude (Max plan): $200/month

Gemini API: $50–$70/month

TinyFish (web agent): ~$50/month

Eleven Labs (voice): ~$50/month

Telegram: free

OpenClaw: open source, free

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

What has truly changed

Dwight saves me 2–3 hours of research every morning. Instead of manually checking X, Hacker News, GitHub trending, and AI blogs, I wake up to a prioritized summary with sources and action items.

Kelly, Pam, and Rachel save me 1–2 hours on content drafting. Ross handles engineering tasks I’d otherwise do at night.

Total: about 4–5 hours saved daily.

But the real value isn’t in any single day—it’s in weeks and months of continuous operation. An agent that researches daily for 30 days accumulates a trackable set of signals, trends, and patterns—something no single session can produce. My posting frequency on X has increased, quality improved, and posting times stabilized. The Awesome LLM Apps repo keeps growing, and the newsletter has a reliable research pipeline.

These agents can’t do original thinking, strategic shifts, or creative breakthroughs. They handle repetitive, structured work that I used to spend hours on, freeing me to focus on tasks that truly require human intelligence.

Getting started

Don’t try to build six agents on day one.

Week 1: One agent, one task. Install OpenClaw, write your first SOUL.md through interaction, pick your most repetitive daily task (research or content drafting for most), set up Telegram, create a cron schedule, and observe it run for a week. Fix issues as they arise.

Week 2: Add memory. Keep refining. Your agent’s initial output will be mediocre—that’s normal. Give feedback, watch the memory grow, and adjust SOUL.md accordingly. By the end of week two, your agent should produce genuinely useful results.

Week 3: Add a second agent. You’ll realize you need a content agent to complement your research agent. Set up shared file workflows: first agent writes, second reads. Coordination is via the filesystem.

Week 4 and beyond: Build sequentially. Add new agents when you feel the pull, not just because you think you should. Each should address a real need—not demos or proofs of concept, but actual gaps in your workflow.

Treat it like hiring. You wouldn’t hire six employees on day one. Hire one, get it running efficiently, then add more as workload demands.

Mindset shift

After your agents run stably for a month, something changes. You no longer see AI as a tool you open when needed, but as a team that’s always working.

I start mornings greeting Monica with a “Good morning,” and before I turn off my phone, I say “Good night” to the team. It sounds silly, but after a month of daily interactions, feedback, and seeing their progress, the line between agent and human begins to blur.

Models are just the baseline—everyone can use Claude, GPT, Gemini. The “system”—built around SOUL.md, memory, scheduling, coordination, and weekly feedback—is yours. No one else has your exact agents, your memory files, or your refined personalities.

And that system compounds daily.

Dwight’s scans enrich his memory; Kelly’s feedback sharpens her drafts; Ross’s bug fixes deepen his understanding of your codebase.

This is the real moat—not the models, but the learning system.

Start today. One agent, one task, one schedule.