Pattern 01 / augmented LLM

Single-Agent Topology: When One Agent Is Enough (2026)

The simplest agent topology. One agent, one model, a curated tool surface, a human supplying the goal. Anthropic’s December 2024 paper calls this foundational unit the augmented LLM.

Single-agent topology. A human supplies a goal. The agent reasons, calls tools, and iterates inside a sense-think-act loop until it produces a result or yields back to the human.Pattern documented in: Anthropic, “Building Effective Agents” (Schluntz, December 2024) anthropic.com/research/building-effective-agents. Accessed 30 April 2026.

The augmented LLM pattern (Anthropic, December 2024)

Anthropic’s engineering essay frames the augmented LLM as the foundational building block. The augmented LLM is a model that can retrieve information, call tools, and remember context across a turn-by-turn loop. The single-agent topology is this building block on its own, without a peer or a supervisor.

The essay’s editorial position is direct: most production deployments do not need multi-agent. The combination of a sensible system prompt, a curated set of tools, and a clear termination condition is sufficient for the majority of structured-task workflows. Where complexity is added, it should be earned by evidence (a measurable failure mode in the simpler shape) rather than chosen by default.

Anthropic’s framing maps cleanly onto an org chart. The single agent is one box; the human is another box; the loop is the closed cycle between them. The tool surface is annotated as a list of available capabilities, not as separate boxes (because tools are not actors with their own goals).

When single-agent is sufficient

Structured information retrieval and synthesis. A research assistant that pulls from a corpus, summarises, and cites is a paradigmatic single-agent task. The work is sequential and well-bounded. There is nothing to parallelise that warrants a peer agent.

Engineering-style edit-test loops. Cognition’s Devin and Anthropic’s claude-code (documented in the public claude-code product description) operate as single agents that read code, edit files, run tests, and iterate. Adding a second peer agent rarely improves the success rate; it usually introduces coordination overhead.

Deterministic-tool workflows. Any workflow where the agent’s decision boundary is “which tool to call next” and the tools themselves are deterministic (a search API, a CRM API, a calculator) is well-served by a single agent. The tools do the heavy lifting; the agent is the dispatcher.

When single-agent is insufficient

Tasks that exceed a single context window. Long-form research, multi-stage code generation across a large codebase, and tasks that require remembering many disparate pieces of state are candidates for supervisor-pattern decomposition or hierarchical decomposition.

Genuinely separable specialised roles. If two roles are best served by different system prompts, tool surfaces, or model sizes, a peer multi-agent or supervisor pattern earns its keep. CrewAI’s researcher + writer + reviewer crew is the canonical illustration.

Reference example

A single-agent engineering loop. An engineer supplies a coding goal. A single agent reads files, edits files, runs tests, executes shell commands, and iterates until tests pass. The shape is documented for Cognition’s Devin and Anthropic’s claude-code.Source: Cognition AI engineering blog, “Introducing Devin” (March 2024) cognition.ai/blog/introducing-devin; Anthropic, claude-code product page anthropic.com/claude-code. Accessed 30 April 2026.

Related on this site

Multi-agent peer topology: when single-agent is no longer enough.
Supervisor pattern: orchestrator and workers.
Human-in-the-loop: oversight when the agent has write access.
Examples gallery: cited single-agent deployments.

For the broader definition of an AI agent, see whatisanaiagent.com. For the engineering reference, see buildingeffectiveagents.com.