Building Effective Agents:
an engineering reference.
An independent reference for engineers building AI agents. Patterns, frameworks, evaluation, failure modes, glossary. Every specific claim links a primary source. No first-person operator voice. No fabricated benchmarks.
Anthropic, in “Building Effective Agents” (Schluntz and Pagnoni, December 2024), distinguishes workflows from agents. A workflow follows a predefined path through code; an agent decides the path at runtime. The same paper names five composable patterns that cover most production agent designs. This site is the engineering reference that extends that paper.
The voice is third-person reference. Vendor names appear as editorial mentions, never as ranked verdicts. The seven editorial rules that govern this site are public on the methodology page.
Sense, think, act, observe.
Modern agent architectures share a four-step loop. The agent senses (reads input and context), thinks (decides the next action), acts (calls a tool or returns output), and observes (reads the result of that action). The loop iterates until a terminal condition is met or a budget cap fires.
The classical formulation is in Russell & Norvig, AIMA, 4th ed. The modern LLM-era formulation reaches the same shape via tool calls and reflection. See agent architecture for the expanded loop with planning, memory, and reflection.
The five patterns.
Five composable patterns from Anthropic's paper. Each essay quotes the original definition, then explains when the pattern is appropriate, where it tends to fail, and which public projects use it.
A linear sequence of LLM calls in which each step's output is the next step's input. Appropriate when a task decomposes cleanly and gating between steps reduces error.
A classifier picks one of N specialised handlers. Appropriate when input classes have meaningfully different cost or quality requirements.
Fan-out to N independent calls, fan-in via aggregator. Two flavours: sectioning (sub-tasks) and voting (same task, multiple attempts).
A central LLM plans, dispatches subtasks to worker LLMs, then merges. Powerful, expensive when the planner over-decomposes.
A generator proposes, an evaluator critiques, the loop repeats until acceptance. Requires a hard cap on iterations.
- An engineering reference, in the shape of a long-form documentation site rather than a blog.
- Citations to primary sources: vendor docs, peer-reviewed research, public benchmarks, vendor pricing pages.
- Custom architecture diagrams, hand-built in line-art SVG, with a consistent visual language across pages.
- Glossary anchors (/glossary/) that other pages and sister sites can deep-link.
- Not a tool review site. Per-tool category overviews appear on /frameworks/ with vendor-published trade-offs and benchmark links, not ranked verdicts.
- Not an operator notebook. No first-person production claims. No fabricated post-mortems, incidents, or simulated logs.
- Not a vendor in disguise. Brand mentions are editorial and fair-use, the same posture as InfoQ or DEV.to.
- Not a moving target without dates. Every page carries a Last verified stamp; refresh discipline is on /methodology/.