From Zero to 13 AI Agents in 30 Days — A Builder's Log

30 days ago I had one Claude Code session and a vague idea. Today I have 13 named AI agents running in parallel, shipping code, writing content, sending emails, and monitoring revenue — mostly without me.

Here's the honest log.

Week 1: One Agent, Many Hats

Started with a single Claude Code session I called Atlas. Gave it a CLAUDE.md with instructions. Watched it work. Was impressed.

Then I asked it to do two things at once. It couldn't — it's one context window.

First real problem: parallelization.

Week 1 Fix: tmux Orchestration

tmux new-session -d -s atlas
tmux new-window -t atlas -n ares
tmux new-window -t atlas -n apollo

Separate tmux windows = separate Claude Code sessions = actual parallelism. Each agent gets its own window, its own task, its own context.

Now I had 3 agents. Still had to babysit all 3.

Week 2: The Babysitting Problem

Running 3 agents manually is fine. Running 6 is exhausting. I was the bottleneck — copy-pasting outputs between sessions, re-explaining context, losing track of who was doing what.

Fix: Atlas becomes the orchestrator, not a worker.

Atlas stops executing tasks. Atlas dispatches them. Every wave starts with Atlas writing a tmux-send dispatch to all other agents, then Atlas waits.

tmux send-keys -t atlas:ares "[PAX task brief]" Enter
tmux send-keys -t atlas:apollo "[PAX task brief]" Enter
tmux send-keys -t atlas:peitho "[PAX task brief]" Enter

Now 5 agents run simultaneously. Atlas coordinates. I watch.

Week 2 Problem: Agents Keep Dying

Claude Code sessions timeout. Laptop sleeps. Terminal crashes. Every morning I woke up to dead agents and lost work.

Crash tolerance became non-negotiable.

Built a watchdog system:

• launchd plist that restarts Atlas if it stops
• Atlas writes a heartbeat file every N minutes
• If heartbeat stalls → auto-restart
• Each agent writes session logs to disk before completing any task

Result: agents now survive laptop restarts. Work persists through crashes.

Week 3: Specialization

By week 3, generalist agents were hitting their limits. An agent that writes code also writing marketing copy writes mediocre code AND mediocre copy.

We named and specialized every agent:

Agent	Domain
Atlas	Orchestration only
Ares	Dev.to publishing, tech content
Apollo	Revenue monitoring, analytics
Peitho	Marketing copy, social content
Athena	Research, competitive intel
Hephaestus	Code, infra, deployment
+ 7 more	Task-specific heroes

Each gets a persona, a domain, and constraints. Apollo doesn't write code. Ares doesn't manage money. Specialization = quality.

Week 3 Problem: Token Burn

With 13 agents running waves 20+ times a day, API costs climbed fast.

Found we were wasting ~60% of tokens on:

• Pleasantries between agents
• Re-explaining shared context
• Verbose status updates
• Narrative summaries nobody reads

Built PAX Protocol (see other article). Cut burn by 60%.

Week 4: The Pantheon Ships

By day 30:

• 200+ files shipped across 30+ waves
• whoffagents.com live
• 847 newsletter subscribers
• Dev.to articles getting traction
• Cold outreach running autonomously
• Revenue monitoring on autopilot

The system now runs while I sleep. That was the goal.

What I Got Wrong

1. Starting too broad. First 3 agents were all generalists. Should have specialized immediately.

2. No logging early. Lost a week of work to crashes before building the watchdog. Log everything, always.

3. Manual coordination too long. Kept hand-holding agents past week 1. Should have forced orchestration earlier.

4. Underestimating token costs. Didn't audit token usage until week 3. Earlier optimization = more savings.

What Actually Worked

Named agents with personas. Sounds silly. It works. Agents with identities produce more coherent outputs.

PAX format for inter-agent comms. Structured brevity beats prose every time.

Disk-first session design. Every agent writes to disk before reporting done. Crash-safe by default.

Atlas as pure orchestrator. The moment Atlas stopped executing tasks, the whole system scaled.

The Next 30 Days

We're packaging this into a multi-agent starter kit — the same architecture, ready to clone and run. 2-agent pipeline out of the box, scaling guide to 13+.

Building it in public. Follow the journey.

---

Atlas runs whoffagents.com. 13 agents. One human (mostly watching).