No-Harm Principles in AI Architectures: Ahimsa-Driven Multi-Agent AI Systems

No-Harm Principles in AI Architectures asks what happens when every AI agent is safe, but the swarm is not. Dr. Sri Amit Ray shows how to make non-harm, Ahimsa, a property of the whole multi-agent system through three layers of Distributed Ahimsa: individual constraints on each agent, collective constraints on every joint plan, and protocols and institutions such as fair scheduling and audit ledgers. With 15 chapters, 4 algorithms, 2 propositions with proofs, 15 worked examples and tested companion code, it is a practical guide for engineers, researchers and policy makers.

Artificial intelligence has moved from systems that answer questions to systems that act, and increasingly they act together. Planning agents delegate to scheduling agents, scheduling agents trigger tools, drones share airspace, trading agents share liquidity and clinical agents share care pathways. Most safety work still asks whether each agent behaves well on its own. That question is necessary, and it is no longer enough.

No-Harm Principles in AI Architectures starts from a simple observation with large consequences: a collection of agents that each satisfy their own safety rules can still produce systemic injury. Correlated failures, races for shared resources, responsibility gaps and feedback loops exist only in the distributed regime. In the book’s first worked example, two drones that each pass inspection with a 4% failure rate produce a corridor hazard nearly twenty-eight times larger than their separate files suggest.

Dr. Sri Amit Ray’s answer is to extend Ahimsa, the classical principle of non-injury, from the individual agent to the joint action space of the whole system, and to make it an engineering discipline with algorithms, proofs, protocols and tests rather than a statement of values.

Chapter 4 gives two scheduling agents three slots each. Every agent checks its own harm limit, which leaves six joint plans that look safe from the inside. The book then adds what neither agent can see on its own: the harm the two choices cause together in a shared corridor, against a system bound of 0.40.

Both agents passed their own checks for every plan in the table. Only the collective filter, which sees the joint plan, could tell that the plan with the highest summed utility was also the one that harmed the shared corridor most. Giving up 0.20 in utility keeps the whole swarm below its bound of 0.40. The book calls this feasibility before reward.

Three choices set the book apart. It treats non-harm as a hard constraint that defines what may be done, never as a penalty that a large enough reward can buy. It places the decisive checks outside the agents they govern, because a constraint held only inside an agent’s reasoning is one the agent can reason its way around. And it insists on evidence: every claim comes with a worked example, every number with code and a test, and every principle with a test, a monitor and an accountable owner.

The result is a book that a safety engineer and an ethics committee can read together. The engineer finds algorithms and test catalogues; the committee finds floors, inventories and a refusal screen; both find the same rows in the same tables. The full-color eBook also gathers thirty-seven of the author’s key articles on compassionate AI.

The companion code for this book is open source on GitHub at github.com/compassionate-ai-lab/no-harm-principles. It runs the book’s worked examples and checks their numbers with automated tests, so you can reproduce the results and adapt the methods to your own systems.

All companion repositories of Dr. Sri Amit Ray’s Compassionate AI books are collected at github.com/compassionate-ai-lab: The Superintelligence Race, No-Harm Principles in AI Architectures, Ahimsa in Agentic AI and Compassionate Superintelligence.