Building Effective Agents
An unusually restrained piece on agent architectures, arguing that most problems are better solved by simple composable patterns than by autonomous agents. The most useful advice in the current agent conversation is about when not to build one.
Guide
AI & Machine Learning
Anthropic
About an hour
Advanced
Free
What it argues
That the word agent covers two different things: workflows, where a model is orchestrated through predefined steps, and agents, where a model directs its own process. Workflows are predictable, cheaper and correct for most problems. Agents are appropriate for a narrower set than the enthusiasm suggests.
The patterns it names
Prompt chaining, routing, parallelisation, orchestrator and worker, and evaluator and optimiser, each with the conditions under which it fits. Naming them makes design discussions substantially clearer.
Why it is worth reading now
The field's default is to reach for the most autonomous option available, which produces systems that are hard to debug, expensive to run and unpredictable in front of users. A well argued case for starting simple is genuinely scarce.
Best for: anyone designing an AI system with more than one model call in it.
Ready to start?
Opens on Anthropic in a new tab.
Stuck on something specific?
Writing only gets you so far. If you want an answer to your situation rather than the general case, book a session and we will work through it together. Every session is free; a few slots open each week.
Follow along
New writing, resources and project ideas land here first.