Three Dharma Questions Before You Ship an AI Agent

Not everything that can be automated should be automated. That sentence should be unnecessary, but in the agent economy the question "Can an agent do this?" is increasingly answered "yes" — and the more important question keeps getting skipped. Before you ship an AI agent, there are three dharma questions worth more than any compliance checklist, and they come from a 2,500-year-old text rather than a Responsible AI policy deck. They map to the Bhagavad Gita's three operational principles: dharma (duty), viveka (discernment), and satya (truth).

I reach for the Gita here for a specific reason: the alternatives have failed. "Move fast and break things" has no concept of duty — and the things it broke were privacy, patient safety, and democratic processes. Effective altruism provided the philosophical scaffolding for one of the largest frauds in financial history, because a framework that says "maximise good outcomes" lets a clever reasoner justify any action as a means to that maximum. And Responsible AI checklists became ethics washing: Google, Microsoft, and Meta each published principles and then dissolved the teams meant to enforce them. Three frameworks, three failure modes — no duty, no process integrity, no discernment.

The Three Questions

One — duty: does this serve the user, or only your efficiency? This is dharma — the obligation to the person your work affects. If the automation exists because it cuts your headcount or lifts your throughput but does nothing for the user, it fails the first test. Builder convenience is not a justification. The user's interest is the anchor.

Two — discernment: if it fails, are the consequences reversible? This is viveka in its most practical form. An agent that recommends a bad restaurant produces a bad meal — reversible, survivable. An agent controlling flight software produces a crash — irreversible, fatal. The question is never "How likely is failure?" Everything fails eventually. The question is "What happens when it does?" Irreversible consequences demand a human in the loop, no matter how confident the system looks.

Three — truth: would you be comfortable if the process were fully transparent? This is satya. If you would not want users, regulators, or the public to see exactly how the system makes its decisions, that is a signal. Not about protecting a trade secret — about the ethics of the decision itself. What cannot bear scrutiny should not operate autonomously.

The Test in Practice

Run Boeing's MCAS through these questions. Did it serve the user? No — it served Boeing's interest in avoiding new certification costs, and the pilots were not even told it existed. Were the consequences reversible? No — it pushed two aircraft into dives that killed 346 people. Would Boeing be comfortable with full transparency? No — the internal emails included an engineer calling the plane "designed by clowns." Every question, a failure. The opposite case is the code quality scanner I built: it serves the user's safety, its outputs are recommendations rather than autonomous actions, and its logic is fully explainable. The difference between the two is not technology. It is discernment.

The vibe coding crisis is a discernment failure by exactly this test — shipping code at machine speed fails the reversibility question, and most teams would fail the transparency question too. These questions are not a brake on building. They are the conditions under which building produces something worthy — and, as I argue in svadharma in AI, trust built this way compounds while exploitation produces debt. When the model itself becomes a commodity, as I explore in when intelligence is free, discernment is the scarce resource left. Chapter 8 of The AI Agent Economy develops the full framework — duty, detached action, and discernment, with the case studies and verse citations a blog post cannot carry. The short version: just because you can ship it does not mean you should.