Recap and final exam
Ten modules to move from a raw model call to a supervised research agent. Here is the course condensed, then the threads that run through it.
The course at a glance
| Module | The essential point |
|---|---|
| 1. What an agent adds | The model chooses the next step; that single property justifies the cost and creates every failure mode |
| 2. The ReAct loop | Thought, action, observation, repeated; three stopping checks — finish, iteration cap, budget |
| 3. Function calling | The description is the specification; validate arguments and turn errors into observations |
| 4. Memory | Working memory is the transcript, cap and summarise; long-term memory is a vector store with TTLs |
| 5. Planning | Plan-and-execute past four tool calls; replan twice at most, preserving completed subgoals |
| 6. Self-critique | Reflection catches shallow errors; a different-model judge and tool-based checks catch the deeper ones |
| 7. Guardrails | Iteration, token and dollar budgets; permission tiers; human confirmation for external side effects |
| 8. Typical failures | Loops, drift, wrong tool, hallucinated arguments, endless replanning, prompt injection through pages |
| 9. Observability | JSON decision log per iteration; replay in three modes; LangGraph once the graph earns it |
| 10. Project | Assembled agent, hand-curated evaluation, ten-item production checklist, incident log |
The threads running through the course
Descriptions are code. Modules 3, 5, 6 and 8 all reduce to the same lever pulled at different heights: the tool description, the plan prompt, the reflection prompt, the injection defense. Every one is a specification the model reads and honours. Prompt engineering is not decoration; it is the API through which you program the agent.
Every unit of autonomy has a matching guardrail. The loop needs a stopping condition; the transcript needs a summariser; the tools need permissions; the plan needs a replan cap; the answer needs verification; the run needs a budget. An agent without the matching guardrail is not a "riskier" agent — it is an incomplete one.
Trust evidence, not confidence. Reflection is the model rating itself and is weak. Tool-based verification consults the world and is strong. The same asymmetry appears everywhere in the course: the trace is stronger than the summary, the evaluation set is stronger than the anecdote, the decision log is stronger than the after-the-fact story of what happened.
Observability is a first-class feature. A non-deterministic system without instrumentation cannot be debugged, cannot be improved, and cannot be trusted. Module 9 is not the last module by accident — it is the last thing that must be true before you open the door to users.
The final exam
The exam has 40 questions covering the ten modules: when an agent is the wrong answer, the three stopping checks of a ReAct loop, writing a tool description a model will call correctly, when to summarise working memory, when to plan and when to replan, the honest limits of self-critique, guardrails against prompt injection through fetched pages, the six failure patterns and their fixes, reading a decision log, and the production checklist.
Several questions present situations to diagnose: a loop that never calls finish, an agent that answers the wrong question after five iterations, an unexpected send_email action triggered by a fetched page, a cost that spikes on a class of questions, a plan that starts over on every replan. It is judgement that is assessed, not the recitation of API surfaces.
On success, your certificate of completion is issued immediately; its number is verifiable by any third party on the platform.
Take the table above and, for each row, ask yourself: "how would I see, from a trace, that this is going wrong?". If you can name the failure signature for loops, for drift, for prompt injection, and if you can describe the ten items on the production checklist without opening the module, you are ready. Good luck!
Final exam
Ready to validate this course?
40 questions drawn at random from the course bank · passing score 70% · verifiable PDF certificate issued immediately on success.
Start the examYou need to be signed in to your InSkillML account with an active subscription. You can also start the exam from My courses.