Agent Algorithmics

Agent Algorithmics treats tool-using AI agents as algorithmic objects rather than as applications of language models. It develops the machinery for operating them under real constraints. Action selection when the budget is finite and shrinking. Verification, abstention, and refusal as control decisions with thresholds you derive rather than tune. Speculation, parallelism, and the protocols that decide when work becomes permanent. Memory, retrieval, capability boundaries, and the evaluation of systems whose output is a trajectory rather than an answer. The argument throughout is that reliable behavior comes from explicit control. A budget the policy can read. A gate whose threshold follows from the cost of being wrong. An abort that is complete at every layer holding state. Language models appear here as components with characterized failure modes, the way a database or a network appears in a systems book. Written for engineers and researchers past the demo stage, who have seen an agent work and now need it to keep working. Sixteen chapters across four parts, 351 pages, 354 references, current through September 2026.