The Smallest Architecture That Can Earn Trust
The opening question of this series was not how to make an AI system appear autonomous. It was how much autonomy a workflow can justify, and what architecture keeps that autonomy useful, observable, and able to stop. The answer is now a conditional one: use an agent where meaningful ambiguity defeats fixed automation, begin with the smallest system that can handle that ambiguity, and add coordination only when measured limits make the next boundary worthwhile. Trust is earned not by maximizing independence but by preserving control as capability grows.
This judgment connects the agent's parts and patterns. A model provides flexible interpretation; tools make data and action explicit; instructions turn capability into a bounded routine; a single-agent loop adapts through observations; orchestration moves control when one context or expertise boundary no longer holds; guardrails, observability, and human handback keep the system governable. None of these is a substitute for the others. A powerful model with weak tools can act on poor context. A well-designed workflow without an exit can persist into harm. A multi-agent system without ownership can distribute confusion faster.
Start with the smallest architecture that fits
The first decision is whether the workflow needs an agent at all. The supplied guides point to nuanced judgment, exceptions, brittle rules, natural-language interpretation, and unstructured information as strong candidate conditions. These conditions describe uncertainty that affects the outcome. They do not describe every large or important workflow. A long sequence of stable transformations may be better implemented as ordinary software, while a short decision may deserve careful agent boundaries if it changes money, access, or a person's case.
The right unit of analysis is the uncertain part of the process. An organization may use fixed software to authenticate a user and retrieve a record, an agent to interpret an unusual request, and a deterministic policy or permission check before an account-changing action. This is often better than granting one agent control over every stage. It keeps autonomy near the judgment that benefits from it and leaves stable, high-consequence transitions explicit.
From there, the smallest credible starting architecture is a bounded single agent. It needs a model capable of the reasoning the task requires, data and action tools with clear interfaces, instructions that describe concrete steps and edge cases, and an evaluation baseline. It also needs a run policy: what counts as completion, which errors can be retried, how many attempts are acceptable, and when a person takes control. The architecture is small because the responsibilities are coherent, not because the prompt is short.
A strong baseline establishes capability before cost optimization. The supplied guidance recommends starting with the most capable available model to understand the task's performance ceiling, then testing smaller or faster models against explicit evaluations. This avoids two opposite mistakes. Using a premium model for every easy retrieval task wastes resources; starting with a weak model can make a tool or instruction problem look like a reasoning limit. The choice should be made at the workflow boundary, where quality, latency, cost, and consequence can be observed together.
The single-agent baseline is appropriate when one domain and one control policy can remain coherent. Add Skills, reusable tools, or prompt templates when the agent needs more defined capability. Use a sequential workflow when stages have real dependencies and traceability matters. Consider routing or specialists when domains require different expertise. Consider parallel work when independent directions can be pursued and their results can be reconciled. Consider evaluator feedback when quality criteria are explicit and another cycle will produce a demonstrable improvement.
The pattern is conditional because the dimensions interact. High-control work tends toward a single agent or a predictable sequence. A broad, open-ended research task may benefit from several focused directions.