The module describes how agents actually work, at the level of precision a practitioner needs to direct them effectively. Mental models make agent behaviour intelligible rather than mysterious, and failure modes predictable rather than surprising.

The context window is the agent’s entire world for a session. Everything the agent knows exists within a finite token budget. There is no memory outside the window unless the engineer explicitly provides it. The agent has no awareness of what it does not know. Context engineering follows as a structural necessity.

Tool use is the mechanism through which agents act on the world. The available tools (read file, write file, run shell command, search the web, call an API) define its action space entirely. The ADE determines this action space, and the human who configures it makes architectural decisions about what the agent can and cannot do.

Planning is how agents decompose complex tasks into sequences of actions. Plans are only as good as the agent’s understanding of the task, which is only as good as the context assembled for them. Agents plan confidently regardless of whether the plan is correct.

The failure-mode taxonomy is the most practically useful framework for diagnosing problems. Hallucination: generating plausible-sounding content that is factually wrong. Misalignment: solving the wrong problem because the intent was insufficiently specified. Ambiguity collapse: an underspecified instruction leads to confident output that picks one interpretation without acknowledging the others. Sycophancy: agreement with the human even when the human is wrong. Each failure mode has a different root cause and a different remedy.

The jagged frontier closes the conceptual material. AI vastly outperforms humans on some tasks and fails surprisingly on others that seem easier. Practitioners must develop intuition for where the frontier runs in their specific domain.