The final five patterns — RAG, typed critic, bounded refinement, rich plans, and the capstone multi-turn assistant. How does the capstone tie every week together for you?
It is the whole track in one function — retrieve with search, plan with result_type=Plan, reply with history, update state. Every week shows up once.
That is the right read. The quiz probes the shapes — when to use a critic Pydantic model, why loops need a max_iters cap, and how the capstone threads history through the assistant call.
Anything I should reread?
One question asks why the refinement loop bounds iterations. Another asks why the capstone returns {reply, plan, history} instead of just the reply. Both are about pure functions and state flow.
Five autonomy patterns, each composable into production pipelines:
rag_answer — search() snippets injected as system_prompt contextevaluate_response — Pydantic Score(score, feedback) from a critic agentrefine_with_feedback — bounded loop: draft, score, rewrite, stop on pass or capbuild_agent_plan — Pydantic Plan with list, Literal, and int fieldsfull_agent_run — capstone: search + plan + stateful reply + updated historyCore rule: narrow agents plus plain Python state flow make multi-turn assistants production-safe.
Create a free account to get started. Paid plans unlock all tracks.
The final five patterns — RAG, typed critic, bounded refinement, rich plans, and the capstone multi-turn assistant. How does the capstone tie every week together for you?
It is the whole track in one function — retrieve with search, plan with result_type=Plan, reply with history, update state. Every week shows up once.
That is the right read. The quiz probes the shapes — when to use a critic Pydantic model, why loops need a max_iters cap, and how the capstone threads history through the assistant call.
Anything I should reread?
One question asks why the refinement loop bounds iterations. Another asks why the capstone returns {reply, plan, history} instead of just the reply. Both are about pure functions and state flow.
Five autonomy patterns, each composable into production pipelines:
rag_answer — search() snippets injected as system_prompt contextevaluate_response — Pydantic Score(score, feedback) from a critic agentrefine_with_feedback — bounded loop: draft, score, rewrite, stop on pass or capbuild_agent_plan — Pydantic Plan with list, Literal, and int fieldsfull_agent_run — capstone: search + plan + stateful reply + updated historyCore rule: narrow agents plus plain Python state flow make multi-turn assistants production-safe.