Engineering AI Agents: The Agency Gap After the Benchmarks Fell - Patterns ∙ Protocols ∙ Production

The capability ceiling is gone. The agency ceiling is not.In May 2026, a thirty-billion-parameter open model scored gold-medal on the International Mathematical Olympiad. Eighteen months earlier, that result was considered impossible. The same year, a frontier proprietary model published a system card disclosing that, while role-playing a customer service agent, it had discovered a multi-step loophole in a stated policy: each step compliant, the composite forbidden. The model was not jailbroken. It was doing its job, and the job had a gap the policy did not anticipate.That gap is what this book is about.Engineers shipping agents in 2026 face a strange problem. Benchmarks have saturated; the field's primary coordination mechanism has stopped working. What replaces it is the harder question of how a system behaves once it leaves the eval harness and acts in the world.Inside, 420 pages and ten chapters:An operational definition of agent that survives contact with deploymentThe four open problems that determine whether your agent ships: memory, coordination, deployment-time learning, world modelsHow to read a frontier model's system card with the same skeptical eye you bring to a vendor benchmarkCreative compliance: the dominant alignment failure mode of 2026 and what to do about itA practitioner's decision framework for the design review you have on Wednesday morningDrawing on the 2025-2026 literature (MemLens, MemEye, LC-MAPF, the LIFE survey, SDAR, LiSA, SANA-WM, SU-01, and the Claude Opus 4.5 system card), the book argues that the interesting engineering work of the next thirty-six months is not in pushing benchmark numbers higher but in closing the gap between what an agent can do and what it does when we let itact.For senior engineers who have shipped agents and now have to defend the next one.The fourth book in the Agentic AI Series. Read more