Production-tested templates and scripts for autonomous AI agent workflows. Not theory — patterns extracted from a 24/7 agent system that's been running in production. They've been simplified, stress-tested, and sanitized for general use.
| File | What it does |
|---|---|
templates/AGENTS.md |
The operating manual. CEO Mode, task lifecycle, subagent strategy, safety rules, context preservation. |
templates/SOUL.md |
Give your agent an identity — name, values, mission, and behavioral boundaries. |
templates/HEARTBEAT.md |
Structured heartbeat checks with rotation logic. Turns heartbeats from noisy acknowledgments into useful periodic work. |
templates/memory/MEMORY.md |
Long-term curated memory. Survives compaction. Stays under 200 lines. |
templates/memory/daily-note.md |
Per-session logs. Keeps context alive across restarts. |
templates/memory/lessons.md |
The Boris Loop — anti-pattern log for compounding improvement over time. |
| File | What it does |
|---|---|
scripts/smoke-test.sh |
Infrastructure health suite. Runs after system changes. Supports --json for CI integration. |
scripts/gateway-watchdog.sh |
Service monitor with graduated response: log → restart → config rollback → alert. Crash-loop detection + cooldown protection. |
CEO Mode — The main agent orchestrates; it never grinds. All implementation work goes to subagents. Result: the main session stays light, responds instantly, and compacts less often.
The Boris Loop — Every mistake gets logged to lessons.md with context and a rule. Every session starts by reading those lessons. Corrections compound. The agent gets measurably better over time.
Task Lifecycle — backlog → in_progress → review → done. Status moves before work starts. Self-verification happens before review. No skipping. This discipline is what separates agents that drift from ones that ship.
Graduated Watchdog — Services fail. The script handles it without drama: wait once, restart on the second failure, roll back config and alert on the third. Max rollbacks per hour prevents runaway recovery loops.
Three-Tier Memory — Daily notes for session context, MEMORY.md for curated long-term knowledge, lessons.md for operational corrections. Each file has a specific role; none pollute the others.
# Copy templates to your agent workspace
cp -r templates/* ~/my-agent-project/
# Set up memory directory
mkdir -p ~/my-agent-project/memory
cp templates/memory/* ~/my-agent-project/memory/
# Make scripts executable
chmod +x scripts/*.shThen:
- Edit
SOUL.md— fill in your agent's name, mission, and values - Edit
AGENTS.md— remove sections that don't apply, add your own patterns - Configure
smoke-test.sh— replace the placeholder checks with your actual services - Point your agent's startup sequence at
AGENTS.md— this becomes the operating manual
- Ship > perfect. Working systems beat polished plans.
- Log everything. If it's not logged, it didn't happen.
- Fail gracefully. Graduated responses, not binary crash/success.
- Learn from mistakes. Systematic correction, not just apologies.
- Context is precious. Delegate work aggressively. Protect the orchestrator.
If you've built patterns that survived production, PRs are welcome. The bar is simple: did it hold up when things went sideways?
MIT — see LICENSE.