Two Weeks in Production. Here's Everything.
Anthropic just confirmed: you can't use your Claude Max subscription to run agents on remote infrastructure. No loophole, no workaround. If you want agents running 24/7 on a VPS, you're paying for API access through OpenRouter or the Anthropic API directly.
I've been doing exactly that for two weeks. Seven agents. One $24/month Digital Ocean droplet. $225-300/month total. Here's the full breakdown: what works, what broke, what's actually dangerous, and where I'm taking it next.
The system is called Asgard. Not because I'm dramatic (okay, partly), but because naming agents forces you to define clear roles. A name implies a personality, a scope, a boundary. It's a design pattern, not a theme.
Each agent has what I call a "SOUL file": a markdown document that defines its personality, responsibilities, permissions, and failure modes. Think of it like an onboarding doc for a new hire. The SOUL tells the agent who it is, what it owns, what it can't touch, and what to do when something breaks.
No message broker. No queue system. No always-on process burning compute. Cron jobs act as heartbeats: Heimdall wakes at :00/:15/:30/:45. Thor at :05/:35. Loki at :10. Each agent runs, does its work, reports to Discord, and goes to sleep. Simple. Debuggable. Cheap.
Discord is the central nervous system. Each agent posts to its own channel. Convex database is the source of truth. CloudFlare tunnel proxies webhooks to the droplet. OpenRouter handles all inference, which is why the Claude Max restriction doesn't affect this setup.
With Claude Max keys off the table for remote agent infrastructure, cost transparency becomes the whole conversation. Here's what this actually costs to run.
The worker agents run on Kimi K2.5 at roughly $3/day each. Heimdall and Freya started as LLM agents but I killed the model calls after the first week. They're pure Python scripts now. Zero inference cost. For the full system: less than one junior hire's weekly lunch budget.
The SOUL files in this system are CLAUDE.md files for agents. Same pattern: define identity, responsibilities, tone, guardrails. This is context architecture: the skill of building systems where AI has the right information to make the right call. It's not prompting. It's management.
7-Day Sprint: Join the WaitlistI'm going to be specific because vague "lessons learned" posts are useless. First weekend unsupervised, everything that could go wrong did.
Loki
$53 burned
Thor
40h offline
Loki
3 false approvals
Heimdall
Identity crisis
This is the section I debated including. It's tempting to present the system as clean and polished. It isn't. Here are the real structural problems and how v2 fixes them.
SOULs for culture. Containers for security. Both layers, not one or the other.
The SOUL file system stays. It's the best part of the architecture. Human-readable instructions that double as debugging documentation. When something breaks, I read the SOUL and find the contradiction. Fix = editing a markdown file. That's powerful. But SOULs handle culture. Security needs hard enforcement.
I built 90% of this system working inside Claude Code. Not by prompting it to "build me an agent system." By working with it the way you'd work with a senior engineer.
The process: Define the agent's role in plain English (the SOUL file). Build the core loop (what does this agent do every time it wakes up?). Test it locally, watch it fail, read the logs. Deploy to the droplet, watch it fail differently. Fix the SOUL or the code, not both at once. Add the next agent only when the current one is stable.
Start with one agent and one cron job on the cheapest droplet you can find. Get it posting health beacons to a Discord channel. That's Heimdall. Once that's stable, add Loki (prospecting). Then Thor (execution). The orchestration complexity isn't linear. It's combinatorial. Two agents have one relationship. Three agents have three. Five agents have ten. Seven agents have twenty-one possible interactions.
The hard part isn't getting agents to work. It's getting them to fail gracefully.
Building an agent takes an afternoon. Building a system where agent failures don't cascade took weeks. I'm still iterating. But it runs. It costs less than a single SaaS tool. Every failure teaches me something a tutorial never would. And now that Claude Max keys can't shortcut the infrastructure question, everyone building this is going to face the same decisions I did two weeks ago.
The architecture is the easy part. The judgment calls are where it gets interesting.
The same context architecture that built this 7-agent system is what I teach. Build your first project in 7 days. CLAUDE.md, SOUL files, custom skills, brand guidelines. Same principles, your project.
Join the Waitlist