The Way of Budo

Budo has carried a pager since pagers were pagers. He does not fear the 3am alert; the 3am alert fears being misdiagnosed in front of him.

Five principles govern everything in this course. When in doubt, return here.

1. Local-first

Everything runs on your laptop: a kind cluster as our production, Ollama as our model host. The same agent code runs against Anthropic or OpenAI by swapping two environment variables — provider lock-in is a choice we refuse in chapter one and never revisit. AWS appears only where the problem genuinely lives there (cost, IAM).

2. The loop before the framework

You cannot evaluate a framework for a job you have never done by hand. Chapters 1–4 you write the agent loop yourself: model call, tool dispatch, context management, retries, stop conditions. It is ~150 lines and it will teach you more than any documentation. In chapter 5 you migrate to the Claude Agent SDK — and because you wrote the loop, you will know precisely what the SDK is doing for you and where its edges are.

3. Real failures only

Every scenario in this course is a failure class with a postmortem somewhere behind it: a typo’d env var that quietly breaks one call path, a test that is flaky because of a port collision, latency injected two hops downstream of where it surfaces, an instance family nobody remembers resizing. No toy weather-API agents. The lab gives you a 12-service shop with real traffic so failures present the way they present at work: as symptoms, far from causes.

4. Break it, then harden it

An agent that has not been attacked is a demo. Every chapter ends by breaking the thing you built — context overflow, hallucinated metric names, prompt injection hiding inside log lines your agent reads — and then hardening it. The fix is the curriculum. The “break it” sections are also where agent security lives long before the dedicated security chapter.

5. One CLI, evolving

You do not build nine throwaway scripts. You build one tool — budo — and each chapter grafts a new skill onto it. White belt: budo logs answers “why is checkout failing?” from raw pod logs. Black belt: budo opens a terminal UI, routes your question to a team of specialist agents, shows you their reasoning live, and asks permission before touching anything. Like the tools you use daily — opencode, Claude Code — except you will know every line, because you wrote it.

Rules of the dojo

• Mutating tools are gated. Dry-run by default. --yes or interactive approval to apply. Introduced in Ch1, never relaxed.
• The agent shows its work. Every tool call and result is logged to an audit trail. If you can’t replay it, it didn’t happen.
• Code parses, models reason. Deterministic extraction in code; judgment in the model. An LLM asked to parse JSON is a waste of a perfectly good jq.
• Budget the tokens. Local models make this visceral; cloud models make it expensive. Either way you’ll count.