My Agent Architecture

By Chad Arimura | Renegade BBS & 8-bit Game | August 11th, 2026

(Note: This piece was authored by a human. I initially let my gtm-agent polish the text, but I ended up reverting almost everything. It took me $\approx 10\times$ longer to undo the "fixes" than it would have taken to just publish it with a few run-on sentences and the number 6 written as a digit instead of a word.)

---

The Objective

A few people have been curious about how I've structured my agents. To provide context, the only project I've publicly discussed is The Daily FM, which was essentially a "scratch my own itch" build.

The broader goal of these experiments is scalability. I want to manage multiple commercial products and a nonprofit organization while keeping the headcount of staff and volunteers significantly lower than traditional models require.

I am roughly one month into this endeavor. While the "AI influencers" are currently screaming about thousands of agents, "swarms," and self-evolving graphs, I'm starting with a modest team of six.

The Agent Roster

My current team acts as a combination of a personal assistant, a calendar manager, and a Linear program manager.

---

Technical Foundation

Runtime & Memory

Currently, I am utilizing Hermes agents, though I expect this to evolve. The architecture is defined by a specific set of files and tools:

• SOUL.md and AGENTS.md (Definitions)
• Custom skills and tools
• MCP servers
• A dedicated Mnemosyne memory bank per profile
• A shared Obsidian wiki (synced locally)

I discovered the "wiki-brain" approach via Andrej Karpathy's LLM dump. While it didn't click immediately, it has become an essential business operating manual, allowing agents to understand "how we do things here" instantly.

The Model Stack

All agents currently run on OpenAI GPT-5.6 Sol, with the ability to spawn sub-agents using GPT-5.6 Terra.

I previously attempted to use Anthropic's Fable, but I encountered two issues:

1. A massive API bill.
2. A lack of subscription-based usage for this specific configuration.

Since frontier models are largely interchangeable for these tasks, I stuck with GPT, as a \100$ subscription is highly cost-effective.

Compute & Security

The entire operation runs on a DigitalOcean Basic Droplet:

• CPU: 4 vCPUs
• RAM: 8 GB (Upgraded from 4 GB to prevent disk swapping)
• Storage: 160 GB Disk

To keep the environment secure, I use Tailscale and ensure zero public ports are exposed.

---

Development & Communication

Coding Workflow

Despite having a dev-agent, I still handle 95% of my coding via Claude Code (with Fable) and Codex. I prefer the tactile experience of the local terminal, where I can observe the reasoning and code changes in real-time.

The Communication Hub: Buzz

I use Buzz, an open-source Slack alternative developed by Block. It is currently in early stages (v0.5.9), requiring a manual iPhone installation, but the agent integration is first-class.

The vps-agent can deploy a new agent and integrate it into Buzz as a team member in about 10 minutes. Most agents can communicate with each other, though the vps-agent is restricted—only I can communicate with it.

The Nostr Protocol

Buzz is built on Nostr, which utilizes "NIPs" (Nostr Implementation Possibilities) to define its specs.

• NIP-01: Defines the core logic: $\text{Universal Signed Event} \rightarrow \text{Key-based Authorship} \rightarrow \text{WebSocket Relays}$.

---

Operational Workflows

I have implemented webhooks that feed external system data directly into Buzz.

Example: Sentry Notification Workflow

# buzz workflows list --channel 7a46....42f
name: sentry_notifier
description: Post Sentry alerts to Buzz
trigger: 
  on: webhook
steps:
  - id: step_1
    name: send_message
    if: trigger_action == "created"
    action: send_message
    text: |
      🚨 Sentry issue created: {{trigger.data | truncate(1000)}} 
      @ops-agent please triage — use your Sentry MCP for full details.

Automation in Practice

1. Sentry Response: An error triggers a Sentry event $\rightarrow$ Buzz Workflow formats the alert $\rightarrow$ ops-agent is tagged. The agent then analyzes Sentry, Cloudflare, and the source code to provide a report and a fix. Current status: Human-in-the-loop.
2. Mobile Dev: If I have an idea while browsing my site on my phone, I message it to Buzz, and an agent implements it.
3. Morning Briefs: ea-agent synthesizes Linear tickets and calendar events to suggest my daily priority list.
4. Social Strategy: gtm-agent reviews the content calendar for gaps. (I remain cautious here; I prefer authentic engagement over "clicks for the sake of clicks").

Future Roadmap

• Establish a full "software factory" pipeline.
• Enable agents to react autonomously to Linear tickets.
• Transition to open-weights models (if incentive outweighs cost).