5 Reasoning: How your agent decides what to do next

The chapter frames reasoning as an architectural control problem, not just “thinking longer.” An agent needs a clear representation of goal, current state, evidence, and constraints so it can produce an inspectable decision artifact rather than an opaque answer. It also stresses that private model reasoning, visible decision records, and action-and-evidence traces are different things: the first may help generation, but only the latter two belong in application-owned governance and audit.

From there, the chapter presents four reusable patterns for deciding how an agent should proceed. Chain-like reasoning makes intermediate claims explicit when dependencies matter; complexity-based routing chooses the right model, effort, or fallback based on difficulty and consequence; parallel exploration preserves genuinely distinct alternatives before committing; and iterative hypothesis testing lets external evidence revise the active explanation. Across all four, the key idea is to match reasoning structure to the task, budget, risk, and available evidence instead of assuming one generic approach fits every request.

The chapter repeatedly warns against mistaking fluent rationales, self-reported confidence, or a long trace for proof. Good design records what was seen, what was tried, what changed, and why a step was accepted or rejected, while keeping recovery authority and consequential actions separate from reasoning itself. In practice, the best system may sometimes use a chain, sometimes route to a faster path, sometimes branch, sometimes probe the world, and sometimes bypass model reasoning entirely in favor of ordinary code and governance.