Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC Berkeley

A model response is not always a finished answer. In the talk's agent-loop example, the surrounding software checks why generation stopped. A request to use a tool means the software must execute that operation and return its result to the model. Reaching the output limit means the response may be incomplete.

That distinction determines whether an automated workflow continues correctly or mistakes unfinished work for success. The model proposes the tool operation; the surrounding program executes it and decides what happens next. Checking the stopping condition is therefore part of building the agent, alongside writing its instructions.

“understanding what you should not do is the key to leading you to what you should do”

Frank Coyle

“The problem is the LLM can't do anything. It is just a probabilistic next word predictor. It can't execute tools.”

Frank Coyle

“If you add the loop, you have Turing computability.”

Frank Coyle

“context means tokens, tokens mean money, and the more context you have, the more confused the LLM is going to be in giving you an answer”

Frank Coyle

“Nothing is a mistake. There's no win and no fail. There's only make.”

Frank Coyle