AI / LLM Integration
Agents: free-recall review
Retrieval beats re-reading. For each prompt, say or write a full answer from memory before you open the model answer — the effort of recall is what makes the agent mental model stick.
Reconstruct the unit’s spine — the ReAct loop, why context cost is quadratic, the several exits a loop needs, the error-recovery trap, and the scripted-vs-agent decision — without looking back at the lesson.
- 01Describe the ReAct agent loop in terms of who does what, and pinpoint what makes it 'agentic'.
- 02Walk a teammate through why an agent's cost grows quadratically with the number of steps, and what overflows at the far end.
- 03Why is a hard max-step cap necessary but not sufficient for safe termination, and what else does a senior add?
- 04Feeding tool errors back to the model is the loop's recovery superpower. How does it become an infinite-loop bug, and how do you bound it?
- 05When should you script a workflow instead of building an open-ended agent, and what does the agent cost you when the path is known?
- 06Name the independent exits a production agent loop should have, and what each one defends against.
If you could reconstruct each answer from memory, you hold the unit’s spine: the agent is a while-loop where the model picks actions and your code runs them; cost is quadratic because the whole transcript is re-sent every turn, and a long run overflows the window and drops its own instructions; a single ‘model stops’ exit is never enough, so you add a step cap, a wall-clock/token budget, and a dedup check; error-feedback is the recovery superpower and the runaway trap, bounded by per-tool retry caps; and you script the known path, reserving the open-ended agent for steps you genuinely cannot enumerate in advance.