Skip to content
WhoWhatWorksBlog

Give AI a Computer: How I Run My Life with Agent Teams

Give AI a Computer: How I Run My Life with Agent Teams

Index

  1. Introduction
  2. The Model Is the Brain. The Harness Is the Body.
  3. What My Setup Looks Like
  4. How I Use It
  5. The Boundaries Matter
  6. Where I Think This Is Going
  7. Start Smaller Than You Think

Introduction

On May 21, 2026, I gave a short AI talk at a VantaSoft happy hour. I showed my coworkers how I use AI for coding, research, project management, and work that needs to keep moving while I’m away from my desk.

The whole talk fit into one sentence:

Give AI a computer. Let it work for you.

I didn’t mean buying a laptop for a chatbot. I meant giving a model a terminal, tools, memory, and a way to talk to me. Instead of opening a new chat and explaining everything again, I can give an agent a task, leave, and check its progress later.

The Model Is the Brain. The Harness Is the Body.

The model handles reasoning. The harness determines what it can actually do.

My harness gives an agent:

  • Access to a terminal and files
  • Tools for taking action
  • Memory across sessions
  • A place to communicate
  • Remote access from my phone
  • Approval limits for sensitive work

Without those pieces, even a great model mostly gives advice. With them, it can work inside an environment I control and return a result.

What My Setup Looks Like

I compared OpenClaw, Hermes Agent, Pi, Dennis’s remote orchestration approach, and Forrest’s Agent Teams. I chose Agent Teams because I understood how it worked and could connect it to tools I already used.

My setup has four main pieces:

Discord

Discord is the remote control. Each agent has a place to receive tasks and report back, and I can check it from my phone.

tmux

Each agent runs in its own tmux session. Closing my terminal doesn’t stop the work underneath it.

Persistent Notes

I use PARA notes and an LLM-maintained wiki for project context, decisions, preferences, and research. This keeps me from repeating the same background every session.

Clear Roles

I use a few agents with specific jobs:

  • Bae handles requests and routes work.
  • Quant focuses on technical and financial research.
  • Mochi tracks goals, priorities, and follow-through.

The names make it sound more complicated than it is. Each agent is a job description with its own context. I only add one when there is a real handoff to own.

How I Use It

Quant watches markets and surfaces things worth reviewing, but it cannot quietly move money. If a decision has real consequences, it asks me first.

Mochi helps manage Oomami by turning loose ideas into tasks and keeping priorities visible. Discord lets me check progress, answer a question, or redirect work while I’m away.

I don’t want zero involvement. I want fewer moments where everything stops because I’m not sitting at my keyboard.

The Boundaries Matter

The biggest benefit is continuity, not full autonomy. My agents remember prior decisions, live in tools I already use, and have roles I can understand when something goes wrong.

I also stay in charge. They can monitor, draft, analyze, and coordinate. I still own the goal, permissions, and final decision.

My original presentation was too confident about account risk. Unofficial tools can create risk when they scrape APIs, reuse credentials, or pool logins, but no setup is completely risk-free. I prefer official authentication, never share personal credentials, check provider terms, and require approval before consequential actions.

This also isn’t the whole topic of agentic engineering. That is about building reliable software with agents, tests, validation, and human judgment. This setup is my personal agent architecture.

Where I Think This Is Going

I expect more agents to communicate with other agents and more routine work to move above traditional interfaces. Instead of clicking through every step, I can state a goal, review the plan, approve sensitive actions, and inspect the result.

Health and finance agents are obvious next ideas for me, but they also need stronger privacy and approval controls. I’m not rushing to give them more access than they need.

Start Smaller Than You Think

I wouldn’t start with a whole team.

  1. Pick one recurring task.
  2. Give one agent a clear job.
  3. Put it somewhere you already use.
  4. Add only the memory it needs.
  5. Require approval for sensitive actions.
  6. Add another agent when there is a real handoff.

My setup became useful before it became impressive. That’s probably why I kept it.

The model is the brain. The harness is the body. Pick one you can keep.

See all posts