Capability and safety are not the same axis.

Autonomous agents now write code, run financial workflows, and take decisions faster than any human can review them. What they cannot yet do is prove they are safe to leave alone.

This book treats that gap as an engineering problem. Ten chapters develop mechanisms that bind agent behavior to a formal or empirical guarantee: rate-distortion orchestration that decides what context is worth passing on, confidence-bounded model checking for workflows that adapt while running, pessimistic offline reinforcement learning that holds up under distribution shift, semantic checkpointing for agents that run for days, trajectory-invariant anomaly detection, drift-aware fairness monitoring, identity-conditioned robust optimization, auditable plan rewriting, and cryptographic binding of the secrets an agent needs in order to act at all.

Each chapter states its problem, develops the method, and reports results on named benchmarks. Written for engineers, researchers, and technology leaders who have to put agents into production and answer for what they do there.