fix(phase-11/17): drop 'primitive' for vendor classes; use class / abstraction / API / loop

phases/11-llm-engineering/17-agent-framework-tradeoffs/docs/en.md

@@ -11,7 +11,7 @@
 
 You have a task that needs more than one LLM call. Maybe it is a research workflow (plan, search, summarize, cite). Maybe it is a code-review pipeline (parse diff, critique, patch, validate). Maybe it is a multi-turn assistant that books flights, writes emails, and files expense reports. You pick a framework.
 
-Three days later, you discover the framework's abstractions leak. CrewAI gives you roles but fights you when the "researcher" needs to hand a structured plan to the "writer." AutoGen gives you chat between agents but has no first-class state so your checkpoint is a pickle of a conversation log. LangGraph gives you a state graph but forces you to name every transition before you know what the agent will do. Agno gives you a single-agent primitive that screams when you try to fan out to three concurrent workers.
+Three days later, you discover the framework's abstractions leak. CrewAI gives you roles but fights you when the "researcher" needs to hand a structured plan to the "writer." AutoGen gives you chat between agents but has no first-class state so your checkpoint is a pickle of a conversation log. LangGraph gives you a state graph but forces you to name every transition before you know what the agent will do. Agno gives you a single-agent abstraction that screams when you try to fan out to three concurrent workers.
 
 The fix is not "pick the best framework." It is to match the framework's core abstraction to the shape of your problem. This lesson draws that map.
 
@@ -97,7 +97,7 @@ Refuse to reach for a framework before you can draw the graph, the org chart, th
 | Research / writing pipeline with distinct roles | CrewAI (sequential) or LangGraph subgraphs | Role-per-task is cheap to express in CrewAI; scale up with LangGraph when branching gets complex. |
 | Proposer-critic or teacher-student dialogue | AutoGen | Two-agent chat is its native shape. |
 | Single agent with tools, sessions, memory | Agno | Thinnest setup, built-in storage and memory. |
-| Thousands of parallel fanouts with reducers | LangGraph + `Send` | The only one with a first-class parallel dispatch primitive. |
+| Thousands of parallel fanouts with reducers | LangGraph + `Send` | The only one with a first-class parallel-dispatch API. |
 | Quick prototype, no framework commitment | Plain Python + provider SDK | No framework is the fastest framework. |
 
 ## Exercises
@@ -117,7 +117,7 @@ Refuse to reach for a framework before you can draw the graph, the org chart, th
 | Crew | "A CrewAI team" | Roles + tasks + process (sequential or hierarchical) bound into a single runnable. |
 | GroupChat | "AutoGen's multi-agent chat" | A managed conversation between N agents with a speaker selector. |
 | Team (Agno) | "Multi-agent Agno" | Route / coordinate / collaborate mode over a set of agents. |
-| StateGraph | "LangGraph's graph" | Typed-state, node, conditional-edge, checkpointer primitive. |
+| StateGraph | "LangGraph's graph" | Typed-state, node, conditional-edge, checkpointer abstraction. |
 
 ## Further Reading
 
@@ -126,7 +126,7 @@ Refuse to reach for a framework before you can draw the graph, the org chart, th
 - [AutoGen documentation](https://microsoft.github.io/autogen/) — ConversableAgent, GroupChat, teams, tools.
 - [Agno documentation](https://docs.agno.com/) — Agent, Team, Workflow, storage, memory.
 - [Anthropic — Building effective agents (Dec 2024)](https://www.anthropic.com/research/building-effective-agents) — pattern library (prompt chaining, routing, parallelization, orchestrator-workers, evaluator-optimizer) framework-agnostic.
-- [Yao et al., "ReAct: Synergizing Reasoning and Acting" (ICLR 2023)](https://arxiv.org/abs/2210.03629) — the primitive every framework dresses up.
+- [Yao et al., "ReAct: Synergizing Reasoning and Acting" (ICLR 2023)](https://arxiv.org/abs/2210.03629) — the loop every framework dresses up.
 - [Wu et al., "AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation" (2023)](https://arxiv.org/abs/2308.08155) — AutoGen's design paper.
 - [Park et al., "Generative Agents: Interactive Simulacra of Human Behavior" (UIST 2023)](https://arxiv.org/abs/2304.03442) — role-play foundation that CrewAI-style persona stacks build on.
 - Phase 11 · 16 (LangGraph) — the framework this lesson benchmarks against.

phases/11-llm-engineering/17-agent-framework-tradeoffs/outputs/skill-framework-picker.md

@@ -12,10 +12,10 @@ Given the task description (problem shape, total LLM calls per run, branching pa
 1. Shape match. One sentence naming the abstraction that fits: graph (typed state, named transitions), org chart (specialist roles, manager-routed handoffs), chat (agents talk until done), single agent with tools. If you cannot pick one, the task is not agent-shaped yet; stop and decompose.
 2. Branching authority. Who picks the next step: developer (explicit edges), manager LLM (CrewAI hierarchical), conversational emergent (AutoGen GroupChat), tool-call self-routed (Agno). Cite the per-turn token cost of LLM-selected routing if applicable.
 3. State budget. Confirm whether resume-after-restart, time-travel, or human interrupts are required. If yes, LangGraph wins on state-first abstractions; Agno covers session-scoped memory only.
-4. Framework choice. Output one of langgraph, crewai, autogen, agno, plain_python. Include the one-sentence justification that maps the shape and state answers onto the framework's core primitive.
+4. Framework choice. Output one of langgraph, crewai, autogen, agno, plain_python. Include the one-sentence justification that maps the shape and state answers onto the framework's core abstraction.
 5. Escape hatch. If the daily run volume is over 10_000 or the task is two or fewer LLM calls without state, recommend plain Python with the provider SDK instead. No framework is the fastest framework when the task is small.
 
-Refuse to recommend AutoGen for deterministic workflows with a known DAG; the GroupChatManager spends tokens picking speakers that the developer could have wired statically. CrewAI does support structured task outputs via `output_pydantic` / `output_json` (see [docs.crewai.com/en/concepts/tasks](https://docs.crewai.com/en/concepts/tasks)), but its `context` channel still flows through the next task's prompt string. Push back on CrewAI when the workflow relies on raw `context` to carry structured state across tasks without one of those output schemas wired up. Push back on LangGraph for a two-call summarizer; the StateGraph overhead is pure tax. Push back on Agno when the task fans out across more than 4 parallel sub-workers with reducer semantics; Agno ships a `Parallel` block whose outputs join into a dict keyed by step name (see [docs-v1.agno.com/workflows_2/overview](https://docs-v1.agno.com/workflows_2/overview) and [docs.agno.com/workflows/access-previous-steps](https://docs.agno.com/workflows/access-previous-steps)), but it does not expose a Send-style fanout-and-reduce primitive comparable to LangGraph's.
+Refuse to recommend AutoGen for deterministic workflows with a known DAG; the GroupChatManager spends tokens picking speakers that the developer could have wired statically. CrewAI does support structured task outputs via `output_pydantic` / `output_json` (see [docs.crewai.com/en/concepts/tasks](https://docs.crewai.com/en/concepts/tasks)), but its `context` channel still flows through the next task's prompt string. Push back on CrewAI when the workflow relies on raw `context` to carry structured state across tasks without one of those output schemas wired up. Push back on LangGraph for a two-call summarizer; the StateGraph overhead is pure tax. Push back on Agno when the task fans out across more than 4 parallel sub-workers with reducer semantics; Agno ships a `Parallel` block whose outputs join into a dict keyed by step name (see [docs-v1.agno.com/workflows_2/overview](https://docs-v1.agno.com/workflows_2/overview) and [docs.agno.com/workflows/access-previous-steps](https://docs.agno.com/workflows/access-previous-steps)), but it does not expose a Send-style fanout-and-reduce API comparable to LangGraph's.
 
 Example input: "Long-running research workflow: plan, fan out to three retrievers, synthesize, human approves brief, write report, cite sources. Must resume after crash. Production-bound to 50 runs per day."
 

phases/11-llm-engineering/17-agent-framework-tradeoffs/quiz.json

@@ -39,7 +39,7 @@
     },
     {
       "stage": "post",
-      "question": "Which framework has built-in storage drivers (SQLite, Postgres, Redis, Mongo, DynamoDB) attached directly to the Agent primitive for session and memory persistence?",
+      "question": "Which framework has built-in storage drivers (SQLite, Postgres, Redis, Mongo, DynamoDB) attached directly to the Agent class for session and memory persistence?",
       "options": [
         "LangGraph",
         "CrewAI",