ModelRefs / Single-Agent System — AI Glossary

Single-Agent System — AI Glossary

An AI system using one LLM as the sole reasoning and action engine, equipped with tools to interact with external environments.

Overview

Single-agent architectures are the simplest agent design: one model receives the system prompt, uses a ReAct or CoT loop, calls tools, and produces a final answer. Sufficient for most tasks within a single LLM's context window. Avoids the coordination overhead and failure modes of multi-agent systems.

Reference details

Topic	agents
Last reviewed	2026-06-24

How it works

One model instance receives the system prompt, the tool schemas, and the task, and holds the entire problem in a single context window.
A loop runs: the model emits either a tool call or a final answer, the harness executes any call and appends the result, and control returns to the model.
All state lives in that growing message list, so there is no separate memory layer to synchronise and no serialisation format between components.
Because every observation stays visible, a decision made at step twenty can be revisited at step forty against the evidence that produced it.
Termination is the model's own judgement that the task is complete, bounded by a step limit the harness enforces from outside.
The binding constraint is the context window, since the loop can only continue while the accumulated transcript still fits alongside the overhead.

Example: The context ceiling is the decomposition trigger

Take a 200,000-token context with a 2,000-token system prompt and 3,000 tokens of tool schemas, so 5,000 tokens of fixed overhead and 195,000 available for work. If a typical cycle — one tool call plus its result plus the model's reasoning about it — consumes 1,200 tokens, the loop sustains 195,000 / 1,200 ≈ 162.5 cycles, meaning 162 complete ones before the window closes. That figure is the honest capacity of the architecture, and it is the number to compare a task against rather than any intuition about complexity. A task needing 40 tool calls has enormous headroom; a task needing 400 does not fit, and no prompt improvement changes that. What decomposition costs is also arithmetic. Splitting across three agents triples the fixed overhead to 3 x 5,000 = 15,000 tokens, which is cheap. The expensive part is the handoff: if each agent compresses 20,000 tokens of accumulated work into a 1,500-token summary for the next, that is a compression of 20,000 / 1,500 ≈ 13.3 to 1, and everything the summary omits is unavailable to every later step. So the single agent's 162 cycles are 162 cycles of full mutual visibility, while a three-agent pipeline buys more total cycles at the price of each stage seeing a thirteenfold reduction of what came before.

Commonly confused with

A single-agent system is not a single model call, and the confusion makes the architecture sound more limited than it is. One call produces one response; a single-agent system runs an unbounded loop of calls against tools, revising its approach as results arrive, and can take hundreds of steps on one task. The distinguishing feature is that one model holds the whole problem, not that it acts once. The second confusion is with multi-agent systems, which are frequently assumed to be the more capable design. On tasks that fit a single context they are usually worse, because the coordination overhead is real and the handoff compression is lossy — the gain arrives only when the task genuinely exceeds what one window can hold, and that threshold is calculable rather than aesthetic.

When to use it

Reach for it when:

For any task whose full working set fits one context window, which the cycle arithmetic above makes a checkable question.
Where steps are interdependent and later decisions need the evidence from earlier ones still visible.
As the default starting architecture, since it has no coordination failure modes to debug.
When latency matters, because there are no handoff round trips between stages.

Reach for something else when:

On tasks that exceed the window, where the loop simply stops rather than degrading gracefully.
Where genuinely parallel subtasks exist, since one agent executes its loop sequentially.
When subtasks need different tools or permissions, which a single tool schema cannot separate.
For workloads needing distinct models per stage, such as a cheap bulk pass feeding an expensive reasoning pass.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Single-Agent System — AI Glossary.

Frequently asked questions

What is Single-Agent System?