Single-Agent System
The simplest thing that deserves the name agent is one model, one loop, one set of tools. A single-agent system is exactly that: a single reasoning process that holds the goal, decides the next action, calls a tool, reads the result, and repeats until the task is done or a limit stops it.
What the architecture looks like
One control loop, one context, one place where state lives. Tools are registered to that agent, memory belongs to it, and every decision is made by the same model with the same view of the task. ReAct is the canonical form — thought, action and observation interleaved in a single trace. Scaling a single-agent system means giving that one agent more: more tools, a longer or better-managed context, sub-routines expressed as tool calls rather than as separate agents.
Why it is usually the right starting point
A multi-agent design buys specialisation and parallelism at a real price — message passing, duplicated work, agents that disagree, and a failure that must be traced across several traces instead of one. A single agent has none of that: one log to read, one prompt to fix, one budget to bound. Its honest limits are a context window that has to hold everything the whole task needs, instructions that get muddled when one prompt tries to cover too many roles, and strictly sequential execution. Reach for more agents when you actually hit those walls, not in anticipation of them.
A single-agent system is an agent built as one reasoning loop with one context and one set of tools: the same model holds the goal, chooses each action, reads each result, and continues until the task completes or a limit stops it. ReAct is its canonical form. It matters as the default architecture — one trace to debug, one budget to bound — with multi-agent designs justified only once that single context or sequential execution becomes the constraint.
Think of one capable person handling a job from start to finish, rather than a team passing it around. They hold the whole goal in mind, pick up whichever tool they need, check the result, and carry on. A single-agent system is software built that way: one decision-maker, one running train of thought, one toolbox.
It is the sensible default. When something goes wrong there is one place to look and one set of instructions to correct. The reason anyone moves to several cooperating programs is that a single worker can only keep so much in mind at once, and can only do one thing at a time.