Side-by-side agent education: each scenario is one SDK-agnostic task under scenarios/<id>/, with idiomatic demos in scenarios/<id>/<framework>/. Compare the same shape across tiers (raw API → agent SDK → platform → managed catalog). Prefer TypeScript when the vendor ships an official TS/JS client. Unit-1 leads with Letta verified cells; employer Managed is secondary. A demo is educationally trustworthy only when scenarios/GAPS.md marks it verified (smoke tripwire green) — linked — pending keys+PASS ≠ verified. Do not invent APIs for gaps (including Decisions API until public docs exist). No Modal.

Sourced from Notion Learning Index + Research docs checks (2026-09-30 PT).

Path convention #

Role	Path
Scenario body	`scenarios/<scenario-id>/README.md`
Per-stack demo	`scenarios/<scenario-id>/<framework>/index.ts` (or `.py`)
Package install + stack notes	`sdks/<framework>/` (this tree)
Coverage matrix	`scenarios/GAPS.md`
Smoke runner	`tripwire/`

cd sdks/letta && npm install
npx tsx ../../scenarios/hello/letta/index.ts
# or:
cd ../../tripwire && npx tsx run.ts --id hello/letta/index.ts

Tier 1 — Raw model APIs (baselines) #

Framework id	Product	Lang	Docs-backed strengths	Honest gaps / lab rules
`claude`	Anthropic Messages	TS + Python	hello, tools, streaming	No native handoffs / HITL / agent memory / MCP harness — leave gap unless a tiny idiomatic pattern exists
`openai`	OpenAI Responses	TS + Python	hello, tools, streaming	Same; Conversations API ≠ agent memory product
`gemini`	Google Gen AI `generateContent`	TS + Python	hello, tools, streaming	Same baseline stance

Tier 2 — Agent SDKs / frameworks #

Framework id	Product	Lang	Docs-backed strengths	Honest gaps / lab rules
`openai-agents`	OpenAI Agents SDK (`@openai/agents`)	TS preferred (+ Python fills)	tools, streaming, sessions, handoffs, approvals, MCP, sandboxes, compaction; computerTool (Phase 2)	Distinct from Agents API hosted `computer_use` and from Decisions API
`claude-agent-sdk`	Claude Agent SDK	TS	query stream, resume, subagents, canUseTool, MCP, FS tools, skills	Not Managed Agents; not raw Messages
`langchain`	LangChain `create_agent`	Python (+ JS streaming)	tools, SqliteSaver memory, HITL middleware, streamEvents	Handoffs only when a tiny subgraph demo is obvious
`eve`	Vercel Eve	TS	filesystem agent, tools, subagents, approvals, skills	Durable memory still a lab gap unless tiny resume demo lands

Tier 3 — Platform runtimes #

Framework id	Product	Lang	Docs-backed strengths	Honest gaps / lab rules
`cloudflare-agents`	Cloudflare Agents (DO)	TS	hello/tools, MemoryAgent `setState`, skills	Streaming consumer + HITL still gap without tiny official quickstart

Tier 4 — Managed / hosted agents #

Unit-1 teaching lead is Letta verified cells (see ../scenarios/UNIT-1.md). Other employer Managed / hosted stacks here are secondary catalog entries — same evidence bar (linked — pending keys+PASS until PASS); do not treat a path as done.

Framework id	Product	Lang	Docs-backed strengths	Honest gaps / lab rules
`letta`	Letta Agent SDK	TS	sessions, stream, resume, tools, approvals, MCP, cloud sandbox	Multi-agent = multiple agents + app orchestration, not `handoffs=[]`. Phase-1 + graduation cells verified where GAPS says so.
`claude-managed`	Claude Managed Agents	TS preferred	Agent/Environment/Session, SSE, memory store, MCP, sandbox, skills	Secondary to Unit-1 Letta/agent-SDK spine. No fake native triage handoff; tool-use may be cloud bash — label in README. Lab cells linked — pending keys+PASS.
`gemini-managed`	Gemini Managed + Interactions	TS preferred	interactions, `previous_interaction_id` + environment, remote env, skills	Secondary catalog. No native handoff; streaming left gap until idiomatic stream consumer is clear; HITL unknown for tiny demos.
`notion-agent`	Notion Agent APIs	TS preferred	sessions, SSE, session continue, `required_actions` approvals	Host-side tool-use stays stub; tools live on Custom Agent / connectors

Index stacks — not yet lab folders #

Per-stack notes: _preview/.

Stack	Status	Lab posture
OpenAI Decisions API	Limited preview; no public API guide	Index-only. Do not invent a Decisions client or folder until official docs ship. Bounded router pattern sits next to handoffs conceptually.
OpenAI Agents API (hosted, beta)	Docs exist; `computer_use` + origin/auth HITL	Prefer SDK `computerTool` under `openai-agents` for Phase 2
Google ADK	Index tutorial track	Folder when Index prioritizes
Bedrock Agents Classic	Closed to new customers	Do not add a Classic folder
Bedrock AgentCore	Platform (any framework)	Phase 3 when Index greenlights; distinct from BMA
Bedrock Managed Agents powered by OpenAI (BMA)	Public preview	Phase 3 separate folder; Python/shell-first until official TS client
Sandboxes: E2B, Daytona, Runloop, Blaxel	Connectors	Pair with a host framework; not Modal

Capability cheat-sheet #

Capability	Strongest frameworks (docs)	Lab note
hello / tools	Nearly all	Notion tools stay stub
streaming	Broad	Prefer dedicated `scenarios/streaming/<fw>/`
memory / resume	letta, claude-agent-sdk, claude-managed, gemini-managed, notion-agent, openai-agents, cloudflare-agents, langchain	Raw APIs = gap
handoffs	openai-agents, eve, claude-agent-sdk	No fake managed handoffs; Decisions ≠ handoffs client
approvals	letta, eve, langchain, openai-agents, claude-agent-sdk, notion-agent	Browser origin/auth gates count
MCP / sandbox-files / skills	Several managed + agent SDKs	Phase 2 fills; computer-use under sandbox + approvals
Decisions (bounded route)	OpenAI Decisions API (preview)	No code until public guide

Learning Index #

Deep self-study tutorials live in Notion Managed Agents: Learning Index (parent of Agent Lab: Curriculum + Capability Matrix). Prefer official quickstart shapes from those tutorials over inventing APIs.

Lab smoke honesty lives in-repo: ../scenarios/GAPS.md — verified vs linked — pending keys+PASS (linked ≠ done). The Notion Curriculum page should mirror that wording (not “filled”).