Building My Agent Team: Larry, Andy, Mary, Hank, and Frankie

The One-Person Team Problem

A solo operator needs marketing, research, compliance, finance, and content production. That is five distinct job functions. You cannot hire five people when the revenue supports one. You cannot ignore any of them either — skip compliance and you lose your license, skip marketing and you lose your customers, skip bookkeeping and April turns into a nightmare.

For eight years running Pacific Cloud Seafoods, I handled all of it myself. Badly, in some cases. The books lived in a Ziploc bag of receipts. Marketing happened when I remembered. Research meant scrolling industry news at midnight. Compliance was the only function that got real attention because regulators do not care about your bandwidth.

AI agents changed this math. Not a single chatbot doing everything. Five agents, each with a defined role, running on my own hardware. A team that does not need health insurance, does not call in sick, and does not require a standup meeting.

Here is who they are and what they do.

Meet the Team

Larry — The Campaign Manager

Larry is an OpenClaw assistant running on a separate machine. His job is marketing workflows — content distribution, campaign management, scheduling, and coordination. Larry manages a Paperclip agent team, which means he is an AI managing other AIs. He decides what gets published where, when outreach happens, and how campaigns are structured.

Running Larry on dedicated hardware was deliberate. Marketing workflows touch external services — social platforms, email tools, analytics. I want that traffic isolated from my primary workstation. If something goes wrong with a marketing integration, it does not take down my development environment or my other agents.

Andy — The Research Assistant

Andy is a NanoClaw running on my main machine inside an Apple container with limited permissions. He finds relevant content, surfaces industry news, and pulls together information I need for writing and decision-making. When I sit down to write a post or evaluate a business opportunity, Andy has already done the legwork.

The sandboxing matters. A research agent needs to access the internet and process external content. That is exactly the kind of workload you want contained. Andy runs with limited permissions inside a container — he can read and summarize, but he cannot modify files outside his sandbox or access other agents' data.

Mary — The Marketing Specialist

Mary is a marketing agent that only runs when I open the app. She handles on-demand work — drafting social posts, reviewing campaign performance, suggesting content angles. Unlike Larry, who runs continuously managing workflows, Mary is the specialist you call into a meeting when you need creative input.

The on-demand model is intentional. Not every agent needs to run 24/7. Mary consumes resources only when I need her. This keeps my machine responsive for the work that actually requires real-time performance.

Hank — The Compliance Veteran

Hank is the HACCP Helper — a voice-enabled inventory and food safety compliance tool. He is the oldest agent on the team, built from years of running a seafood processing operation. Hank speaks fish industry language. You tell him you received 200 pounds of sockeye from Bristol Bay at 33 degrees and he logs it correctly — species, weight, origin, temperature, supplier.

Hank handles receiving logs, temperature checks, and compliance documentation hands-free. In a 34-degree processing room with gloves on, you cannot type. But you can talk. Hank understands the difference between "reds" and "sockeye" because they are the same fish and anyone who works the industry knows that.

He is the agent I am most confident in because he was built from direct operational pain, not theoretical need.

Frankie — The Bookkeeper

Frankie is the finance agent — Frankie-Ledger. She sorts receipts, categorizes expenses, and keeps bookkeeping manageable. Frankie was born from the Ziploc bag era. When your filing system is a gallon bag stuffed with crumpled receipts, and your accountant sends you that look every March, you build a better system or you suffer.

Frankie does not replace an accountant. She makes sure the accountant has clean data to work with. Receipts get categorized when they come in, not in a panic the week before the deadline. Expenses are tagged by category and business function. The result is a bookkeeping workflow that stays current instead of one that requires an annual archaeological dig.

Why Five Agents, Not One

The obvious question: why not one agent that does everything?

Containment. Each agent has a defined scope and limited permissions. Andy runs in a sandbox. Larry runs on separate hardware. Mary only activates on demand. This is not about convenience — it is about security and reliability.

A single mega-agent with access to your finances, your marketing accounts, your compliance records, and your research tools is a single point of failure with an enormous attack surface. If it hallucinates, it hallucinates everywhere. If it gets compromised, everything is compromised.

Separate agents mean separate failure domains. If Frankie miscategorizes an expense, it does not affect Hank's compliance logs. If Larry's marketing workflow breaks, Andy still delivers research. You can fix one agent without touching the others.

This mirrors how you would manage human employees. You would not hire one person to do marketing, accounting, research, compliance, and content. You would hire specialists with clear responsibilities. Same principle applies to AI agents.

Managing Agents Like Managing People

The agents handle execution. I handle judgment. This distinction matters.

Larry can distribute content across channels, but I decide what message we are sending and whether it represents the business accurately. Andy can surface ten relevant articles, but I decide which ones matter and what conclusions to draw. Frankie can categorize a receipt, but I decide whether that expense was justified.

You still review outputs. You still set priorities. You still catch mistakes. The feedback loop is faster than with human employees because agents do not have feelings about corrections and they apply changes immediately. But the loop still exists. Removing the human from the loop is not the goal. Removing the tedium from the human is the goal.

The biggest management challenge is context. Right now, my agents do not talk to each other. Larry does not know what Andy researched. Frankie does not know what Mary recommended for marketing spend. I am the message bus. Every piece of context that needs to move between agents passes through me manually.

This is the single biggest limitation of the current setup.

What Is Working

Hank saves 45 minutes a day in processing environments. That number is real and consistent.

Frankie eliminated the annual receipt panic. Bookkeeping stays current within a week instead of falling six months behind.

Andy cuts research time by roughly 60 percent. Industry news, competitor analysis, and content research that took two hours now takes forty-five minutes.

Larry manages distribution workflows I simply would not do manually. Before Larry, marketing happened sporadically. Now it happens on a schedule because Larry does not forget and does not get busy with something else.

Mary provides creative input on demand without the overhead of a standing meeting or a retainer with a marketing consultant.

Collectively, these agents give a one-person operation capabilities that previously required three to four people. Not at the same quality as dedicated humans in every case — but at a level that is good enough to compete and consistent enough to rely on.

What Is Not Working Yet

Agent coordination is manual. I mentioned this already but it deserves emphasis. The lack of shared context between agents creates real friction. When I want Andy's research to inform Larry's marketing strategy, I have to copy information between systems. That is the opposite of automation.

Quality varies by task complexity. For routine, well-defined work — categorizing receipts, logging inventory, distributing scheduled content — the agents are reliable. For nuanced judgment calls — is this the right marketing angle, does this research conclusion hold up — they need heavy oversight.

Error detection is on me. If Frankie miscategorizes an expense, nothing flags it automatically. I catch it during review or I do not catch it. There is no agent watching the other agents. Yet.

Setup and maintenance is real work. Each agent required significant configuration. Keeping them updated, adjusting prompts, fixing edge cases — this is ongoing operational overhead. It is less than managing human employees, but it is not zero.

The Convergence

The current setup works but it is not the end state. The vision is consolidation — all five agents running locally on my own hardware, sharing context through a unified memory layer, coordinated without me acting as the message bus.

Self-sufficiency drives this. I do not want my business operations dependent on cloud services that can change pricing, alter APIs, or go offline. Running locally means independence from internet outages, vendor decisions, and subscription models that extract more value than they create. This is the same instinct that makes someone keep a generator and a chest freezer. Preparedness is not paranoia — it is operational planning for when things go sideways.

The technical pieces are coming together. Local language models are getting capable enough. Vector databases run on consumer hardware. The memory layer that lets agents share context is something I am actively building. The goal is a system where Andy's research automatically informs Larry's campaigns, where Frankie's expense data feeds into Mary's budget recommendations, where Hank's compliance records are available to any agent that needs them.

A one-person operation that punches above its weight. Not because the operator works harder, but because the team never clocks out and the infrastructure answers to no one but the operator.

That is the future I am building toward. One agent at a time.