Agents
When to build an agent, how the loop works, and how to keep it from going off the rails.
7 topics
Agent or Workflow?
The first decision in any LLM system — whether the model controls the flow, or your code does.
core20 minThe Agent Loop
Every agent is the same twenty-line while loop — model, tool call, result, repeat — and all the difficulty is in the termination conditions.
core20 minDesigning an Agent's Tool Surface
The tools you expose are the agent's entire vocabulary — and choosing between one general tool and twenty specific ones is the highest-leverage design call you make.
core20 minAgent Memory & State
The context window is working memory that vanishes; anything an agent must remember across a long run or a later session has to live somewhere else.
core20 minHuman in the Loop & Approvals
Deciding which actions an agent may take alone, which need a human, and how to ask without making the product unusable.
core20 minMulti-Agent Systems
Splitting work across several agents buys parallelism and context isolation, and costs you coordination, tokens, and most of your ability to debug.
deep20 minDebugging & Observing Agents
You cannot debug an agent by reading its final answer — you need the full trace of every turn, tool call, and result, and the failure is usually in the context.
core20 min