Part VI: Orchestration: Agents, Retrieval, Context

"reasoning traces help the model induce, track, and update action plans."

Shunyu Yao et al., "ReAct: Synergizing Reasoning and Acting in Language Models"

Once a model can be served, the next temptation is to give it work that does not fit inside one answer. It must read, remember, retrieve, call tools, wait for results, revise its state, and sometimes coordinate with other agents. This part is about the machinery around the model when the unit of work becomes a task rather than a completion.

Chapter 37 begins with the training problem: an acting system receives reward over trajectories, one trajectory being a full run of the agent acting step by step, not just answers. Chapter 38 and Chapter 41 then describe the execution loop and the runtime that keeps it from becoming an unbounded script, with Chapter 39 between them separating logs, workspaces, and long-term stores, and Chapter 40 following the memory that outlives any task: the product's memory of its user, and what it does to markets, privacy, and sycophancy. Chapter 42 extends the action space to the screen itself, where the agent operates software built for human eyes and hands. Chapter 43 asks when more than one model call gives a useful signal rather than duplicated failure. Outside knowledge has to be found, ranked, embedded, and treated as untrusted input, and Chapter 44 and Chapter 45 explain how. Chapter 46 closes the part by treating the context window as a programmed surface, with budget, placement, compression, and recursion.

Clean primitives are scarce here. Orchestration is where state leaks, tools hang, memory becomes evidence and liability, and the boundary between model behavior and system behavior blurs. The useful question is not whether an agent is intelligent. It is which loop owns the next decision, what state it can see, and what can stop it when the world disagrees.