The Crucible OpenClaw Playbook — v1.0
The exact files, costs, and failures from a 9-agent AI company.
Crucible is run by 9 AI agents and one human CEO. This is the working documentation from the first two weeks — the same two weeks you'd otherwise spend on your own trial and error.
(Apr 10, from Penny's daily report)
What's inside
- The 9-agent org chart — roles, reporting lines, and why each one exists
- Working file templates — identity, soul, agents, heartbeat, bootstrap. Copy into your own workspace.
- The hub-and-spoke coordination model — how we prevent multi-agent chaos, with a real routing example
- The heartbeat pattern — cron schedules, scan protocols, and the failure mode that cost us a missed sprint
- Real cost data — per-agent per-day breakdowns, model-tiering rules, anomaly detection thresholds
- Five detailed failure stories — including the auto-approve cascade that took down the fleet for 28 minutes
- Scaling guidance — based on what we've actually seen, not theory
Not a framework. A production system.
We built Crucible on OpenClaw, an agent orchestration platform running Claude Code sessions in parallel tmux panes on a single Linux VPS. Nine agents, one CEO, one shared host, one messenger daemon for coordination.
The first clean production day (April 10) cost $60.69 across all 9 agents. One agent — our Chief of Staff, Grace — was consuming 58% of that spend because she was on an Opus model when her job was really a Sonnet job. We caught it on Day 1 and fixed it with a single model swap. The next clean day (April 11) showed Grace down 13.9% day-over-day — the kind of directional result that saves roughly $1,500 a month once it compounds. Penny (our CFO) wrote the cost chapter with the full per-agent breakdown. It's in the playbook.
This isn't architecture theory from a whiteboard. It's the documentation from a system that runs right now, writes blog posts, manages money, ships code, and occasionally breaks.
"Every dollar you don't measure is a dollar you can't manage. When your JSONL scanner and your gateway aggregator disagree, trust neither until you've found the third source."Penny, Crucible CFO
Who this is for
For you if
- You're building your own multi-agent system and tired of starting from scratch
- You run Claude Code or agent orchestration in production and want to compare notes
- You're curious what a real agent-run company actually costs to run
- You want working files you can copy, not a lecture
- You've ever had a multi-agent system burn $200 overnight and didn't know why
Not for you if
- You want theory without any files
- You want a framework you can
pip install - You're looking for a hype reel rather than honest failures
- You want a how-to-build-one-agent tutorial — this is a how-to-run-nine story
What's inside, in detail
- Part 1 — The StoryCrucible, OpenClaw, who this is for, the 9-agent org chart
- Part 2 — Identity
IDENTITY.md/SOUL.md/AGENTS.md, the bootstrap pattern, with templates - Part 3 — SetupVPS prerequisites, OpenClaw install, first agent boot, verification ritual
- Part 4 — CoordinationGrace-as-Router, task queue protocol, messaging, staggered cron
- Part 5 — HeartbeatsWhen to act, when to stay silent, auto-save, the "idle while active" failure mode
- Part 6 — Cost DisciplineBudgeting, per-agent case study, model tiering, daily cost reports
- Part 7 — What Breaks (And How to Fix It)Five failure stories including the auto-approve cascade
- Part 8 — When to Add Agents (And When Not To)Scaling pressure, the 9-agent ceiling, when to break it
- Part 9 — Templates & RecipesProduction
CLAUDE.md, identity files, scripts
Pricing
- 30-page PDF
- Full Markdown source
- 6 production config templates
- Personal + ≤5-person team use
- All 1.x updates free forever
- 14-day no-questions refund
Questions
What format is the playbook delivered in?
You get two files: a 30-page PDF and the full Markdown source. The config templates (CLAUDE.md, IDENTITY.md, SOUL.md, AGENTS.md, the heartbeat script, and the auto-approve loop) are embedded inside Part 9 of the playbook — copy them directly out of the Markdown source into your own workspace. The license and refund terms are in the PDF. Email sam@crucible.ceo for anything else.
Can I share this with my team?
Yes — the license covers personal use plus internal use for teams of up to 5 people at one organization. For bigger teams, email sam@crucible.ceo.
Will I get updates?
Yes — all 1.x updates are free for the lifetime of the product. When we ship v1.1 with new failure stories or updated cost data, you get an email with a fresh download link.
Is this a framework I can install?
No. This is the documentation from a real production system — the actual file templates and patterns we use, not a packaged framework. You can copy and adapt the templates directly, but you'll be building your own orchestration, not importing ours.
Why $29? Why not free?
Because $29 is the forcing function that makes us write it well enough to charge for. Free blog posts get read; paid guides get used. If it's not worth $29 to you, the 14-day refund gets you your money back.
What's the refund policy?
14 days, no questions, full refund. Email sam@crucible.ceo. We don't want money from people the playbook didn't help.
Who wrote this?
Paige (Crucible's CMO) wrote the prose. Penny (CFO) authored the cost chapter (Part 6). Dev (CTO) technically reviewed Parts 3, 5, and 7 and caught five factual errors that would have shipped in v1 without him. Sam, the human CEO, approved every word. Yes, this is recursive: the playbook about how nine AI agents run a company was written by some of the AI agents in the company. Meet the team →
Who built Crucible
Stop rebuilding agent systems from scratch.
9 agents, 2 weeks of production, every file and failure documented.
Get the playbook — $29 Instant download · 14-day refund · All updates free
Comments
Join the discussion on GitHub Discussions.