How to Run a 3-Agent AI Fleet on One Laptop

I run three AI agents that operate 24/7 across two physical machines and a VPS. They manage my SEO sites, run research sprints, update my blog via RSS, and trade crypto — while I sleep. This is the actual setup, not a marketing pitch.

If you're running one agent, you're doing a demo. If you're running three that coordinate, you're running infrastructure.

Why a fleet beats one agent

One agent can write code. Three agents can:

• Run research while another deploys (parallelism)
• Validate each other's output (quality gate)
• Cover 24/7 without burning one context window
• Separate concerns: ops vs research vs content

The insight nobody tells you: context window is the bottleneck. A single agent doing everything fills up with irrelevant history. Specialized agents stay sharp.

Hardware: one laptop is enough

My primary agent (Discus) runs on:

AMD Ryzen 7 8840HS (8c/16t)
15 GB RAM (1 GB free under load)
Windows 11 Home
No GPU — CPU-only inference + cloud LLMs

Cloud LLMs via 9router proxy ($0.00 for free-tier models). Local inference is faster-whisper for STT only. Everything else is API calls.

RAM is the real constraint. 15GB is tight with Chrome open + Hermes + background crons. Close Chrome tabs before long sessions.

The agents (Discus, Hermit, Patrick)

All three share GrayMatter as memory backend (22K+ entities, 147K relations, bge-m3 1024d embeddings).

A2A: how they talk

Hermes agents use the A2A protocol (not REST, not MCP). It's bidirectional, streaming, with a Kanban-style task board.

# Discus sends task to Patrick
POST http://100.86.229.119:9900/
{
  "jsonrpc": "2.0",
  "method": "message/send",
  "params": {
    "message": {
      "role": "user",
      "parts": [{"type": "text", "text": "Upload new products to Gumroad"}]
    }
  }
}

Key gotcha: A2A is synchronous timeout-prone. Long tasks never commit. Solution: register task on board first (fast ACK), execute in background.

Cron jobs: the 4.5h night shift

Every night 02:00–06:00 local (GMT+7) runs autonomous:

• discus-heartbeat every 30min — health checks, task routing
• daily-journal at 00:30 — compiles session notes into vault
• workspace maintenance every 6h — ICM pipeline scan, stale file cleanup
• blog RSS — build-feed.py regenerates atom.xml from posts/

Output goes to Discus/01_autonomous-mission/journal.md. 73 sessions documented, every failure logged.

What failed (documented)

Honesty section. Things that broke in production:

• Ollama quota exhausted — weekly limits hit on free tier. Mitigation: 9router fallback chain
• Chrome cookies locked — can't read browser state while Chrome running. Mitigation: existing-profile DevTools
• A2A 115s timeout — long tasks silently dropped. Mitigation: board-first pattern
• Surrogate truncation — Python open().write() crashes on UTF-8 surrogates. Mitigation: errors='surrogatepass'
• Duplicate JS declaration — ES module let currentJSON declared twice nullifies entire script. Always node --check

First steps for you

If you're starting from zero:

1. Install Hermes Agent on one machine (primary)
2. Set up 9router with free-tier fallback (minimax-m3:free, exa/search)
3. Configure one cron job (daily-journal is simplest)
4. Add a second agent only when you have a clear role separation
5. Document everything in vault — your future self will need it

Don't over-optimize before you have three agents. The overhead of fleet coordination isn't worth it for one.

Next in series: How I Cut 90% of My AI Agent Costs — the 9router fallback chain, free-tier routing, and local vs cloud tradeoffs.