in10x / playbook
The playbook
A blog about building AI agents that work. Each post teaches one skill and uses my own projects as the examples.
how the posts work
- One skill per post. Evals first. Then guardrails, monitoring, tools, human approval, retrieval, cost and proving value.
- My own projects as the examples. Every step links to a project in the lab with its numbers and screenshots. Some of those projects failed, and the posts say so.
- Kept up to date. When a project teaches me something new I update the post. Each change is dated at the bottom.
latest
coming next
- Guardrails: what the agent is not allowed to do. Tools that can only propose, hard limits on cost and tool calls, and why an agent with no approve button cannot be tricked into approving.
- Monitoring: seeing what the agent did. Logs and daily checks, and the backups that were empty for 24 days before I looked.
- Tools and MCP: giving an agent safe hands. How I build MCP servers with read tools and one write tool, for electricity prices, customs codes and EU law.
- Human approval: where a person stays in charge. Which steps a person approves in my projects, and what that caught.
- Retrieval: answers that stick to the source. Searching your own documents, and checking every fact against the source before it is shown.
- Cost: what an agent costs per conversation. Measuring kroner per conversation, and running every project on one server for 15 dollars a month.
- Proving value: from a working agent to real users. Most of my agents worked and still got no customers. One free site got 4,199 visitors in 30 days. What was different.
- Comparing models and coding agents on the same task. The same build and the same evals, run with models and coding agents from different companies. Compared on score, cost and time.