🧠 Notes on building effective agents
Based on Building effective agents from the Engineering team at Anthropic
Use simple, composable patterns (like lego bricks)
Anthropic distinction between agentic systems:
A. Workflows - predefined code paths
B. Agents - LLM dynamically direct their goal
When working with LLMs, find the simplest solution possible.
Often, an agentic system solution is not necessary: optimizing a single LLM call with retrivial and in-context examples is enough.
About using agentic frameworks such as langraph
They make it easy to get started, simplifying low-level tasks such as calling LLMs, defining and parsing tools and chaining calls together.
They obfuscate the prompts, making it hard to debug.
Anthropic suggest to start using LLMs API call directly.
Building block:
Augmented LLM
Workflow: prompt chaining
Workflow: routing
Workflow: parallalization
Workflow: orchestrator / workers
Workflow: evaluator-optimzer
While building tools:
Put yourself in the model's shoes
function main(){
const x = "y"
}