BuildingEffectiveAgents
title: Building Effective Agents (Anthropic) radar_quadrant: Techniques radar_ring: Assess radar_position: inner created: 2026-05-26 last_updated: 2026-05-26 tags: [llm, agents, workflows, tool-use, anthropic, architecture] source_url: https://www.anthropic.com/research/building-effective-agents
A design guide published by Anthropic's research team covering the architecture of LLM-based agent systems. Widely cited as a primary reference for teams deciding how to structure LLM workflows and autonomous agents.
The guide argues that most production use cases are better served by simple, well-defined workflows than by fully autonomous agents. Complexity and autonomy should be added only when simpler approaches are insufficient. The framing shifts the design question from "how do I build an agent?" to "does this problem actually require an agent?"
The guide defines a spectrum of patterns in increasing order of autonomy:
- Prompt chaining: sequential LLM calls where each output feeds the next input; low autonomy, high predictability.
- Routing: a classifier directs inputs to specialised subprompts or models.
- Parallelisation: independent subtasks run concurrently and results are aggregated.
- Orchestrator-subagents: a central LLM breaks tasks into subtasks and delegates to worker agents.
- Evaluator-optimiser: one LLM generates outputs, another scores and refines them in a loop.
Agents are defined as systems where an LLM dynamically selects tools and determines its own action sequence. The guide identifies the conditions that justify agents: tasks with ambiguous steps that cannot be predetermined, multi-step reasoning over external state, and situations where human oversight at each step is impractical. It also covers failure modes: tool misuse, runaway loops, and compounding errors in long chains.
Complements Patterns for Building LLM-based Systems & Products (Eugene Yan), which covers a broader taxonomy. This guide is narrower and more authoritative on the agent/workflow boundary specifically. Also complements AutoGen Magentic-One (multi-agent orchestration platform) and LLM Evaluation Methodology.
Placed in Techniques / Assess / inner. High-authority canonical reference from the team that builds the models. The workflow-vs-agent decision framework is directly applicable before any LLM system design. Inner position reflects zero tooling cost — applying the framework requires only a design conversation. Trial gate: one system designed with explicit reference to the workflow/agent taxonomy, with the chosen pattern documented and justified against the guide's criteria.
