Zen and the Art of Agent Engineering
Early this month, I was in the Himalayas for 冈仁波齐转山: the kora around Mount Kailash.
Sacred to four religions. Walked clockwise by pilgrims for over a thousand years. Not supposed to summit.
Day one at altitude. Gasping for breath, thinking about agents. What we face had older names already.
缘起: Dependent Origination
Nothing arises in isolation. There is no "the agent": only this model in the harness, with prompt, tools, memory. Change any condition, behavior changes.
Context engineering, harness design: these are just dressed-up words for arranging the conditions deliberately.
无明: Fundamental Ignorance
Most engineers treat hallucination as a bug. It isn't. It is the model's ground state: a system trained to predict what sounds right has no way to tell "I know this" from "this sounds like something I'd know."
More parameters won't fix this. It is the system seeing its own limits: calibrated uncertainty, grounded retrieval, honest refusal.
无常: Impermanence
What worked yesterday might not work tomorrow.
The Sword of Damocles hangs over every engineer in 2026. We are worried about being outpaced. And yet we choose to learn, build, and ship with LLMs every day. 成住坏空 (arising, abiding, changing, ceasing) has come for our profession. We just happen to be living through the arising.
戒定慧: The Three Trainings
- 戒 Discipline: What the agent cannot do. Auth gateway, policy engine, kill switch.
- 定 Concentration: Single purpose, narrow scope. Whatever tries to do everything does nothing well.
- 慧 Wisdom: Knowing when the sandbox itself needs to change, not just the contents. The hardest layer to automate. Without it, you build the wrong thing well.
· · ·
Summiting 冈仁波齐 was never the right verb.
The trail is slope after slope. The hardest part isn't the slope you're climbing. It's the next slope, already visible up ahead, while you're still gasping through this one.
Agent engineering is a kora. Every slope teaches what the last one couldn't.
The practice is the work.