Workflow Graphs & Agent Frameworks
How do you draw multiple steps as a work map?
Choose between workflows, agents, graphs, and frameworks.
📌 Learning goals
- Explain the difference between an agent loop, a workflow graph, an agent framework, and a multi-role system in your own words.
- Choose the simplest tool that can finish the task instead of adding roles just because they are fashionable.
- Complete five exercises and compare LangGraph, CrewAI, Smolagents, and Pydantic AI hands-on.
- Explain which problems handoff, checkpointing, and human approval each solve.
Entry conditions
Finish the six Stage 3 exercises and be able to name schema → call → execute → result → answer; reading async/await helps but is not required to start.
🧭 Lessons on this site
Read in the suggested order; checkboxes share the same browser progress as the /learn track pages.- 01Complete LangChain Tutorial 2026: Building Enterprise-Grade LLM Applications from Scratch
The most detailed LangChain 2026 tutorial: covering LCEL, RAG, Agent, LangGraph, LangSmith. From installation to production deployment, with complete code examples.
53 min - 02LangChain vs LlamaIndex: A Selection Guide and Hands-On Primer
A deep comparison of two major LLM application frameworks: LangChain's generality vs LlamaIndex's data indexing expertise, with starter code and scenario recommendations.
9 min - 03Three-Layer Agent Collaboration Framework: Hermes → OpenClaw → Claude Code Architecture Explained
An in-depth analysis of a three-layer AI Agent collaboration architecture built on Hermes Agent, OpenClaw, and Claude Code: the responsibilities of each layer, the communication mechanism, and why it is designed this way.
6 min - 04A Complete Tutorial for Open Multi-Agent (OMA): A TypeScript Multi-Agent Orchestration Framework from Goal to Task DAG
An in-depth look at OMA (open-multi-agent) v1.8.0: a TypeScript-native multi-Agent orchestration framework with Goal-Driven Task DAG, Checkpoint resume, Consensus verification, and support for 10+ LLM providers. Includes complete code examples and a hands-on tutorial.
43 min - 05The Evolution of the Harness: Six Eras, Each With Its New Capabilities and New Ways to Crash
From pure-prompt chatbots, CoT and ReAct reasoning prompts, and function calling — the real turning point — to MCP standardization, long-running CLI Agents, and always-on multi-Agent orchestration: era by era, what new capability the harness gained each time, and what new failure mode that capability incidentally introduced.
12 min - 06A Multi-Agent Investment-Research Pipeline: Collect, Analyze, Report in Three Stages
A single Agent cannot finish an investment-research task: the context blows up, attention scatters, and errors cross-contaminate. This article gives the complete design for a three-stage collect / analyze / report division of labor — each stage's responsibilities and negative list, the structured JSON contract between stages, the write-scope partition, failure isolation and contamination detection, where to place the verification gates, and the cost-and-latency trade-off framework, with an architecture diagram and a failure-mode list.
23 min
📚 Required reading
- 1.Anthropic — Building Effective Agents⭐⭐⭐⭐⭐Separate workflows from agents, and see why you should start with the simplest option.
- 2.LangGraph — Workflows and Agents⭐⭐⭐⭐⭐See how fixed and dynamic routes are written as graphs.
- 3.OpenAI Agents SDK — Multi-agent orchestration⭐⭐⭐⭐Compare manager-as-tools with handoffs.
🎯 Curated resources
| Resource | Who it's for | Priority | Why |
|---|---|---|---|
Production frameworks langchain-ai/langgraph | Need state, checkpoints, HITL, and replayable flows | ⭐⭐⭐⭐⭐ | A low-level runtime; state, edges, checkpoints, and interrupts stay visible, at the cost of more design work. |
Production frameworks OpenAI Agents SDK | Already in the OpenAI ecosystem | ⭐⭐⭐⭐⭐ | Handoffs, guardrails, and tracing; Sandbox Agents are beta, not a solved production story. |
Fast prototyping / multi-agent crewAIInc/crewAI | Want role flows fast | ⭐⭐⭐⭐ | Spin up researcher → writer → reviewer flows fast; Flows now support persistence and human feedback. |
Special routes Hugging Face Smolagents | Comparing CodeAct with tool calling | ⭐⭐⭐⭐ | Model-generated code must run in isolation. |
Special routes pydantic/pydantic-ai | Care about typed, structured output | ⭐⭐⭐ | The schema validates shape, not semantic truth. |
Special routes langchain-ai/deepagents | Want a full harness with planning, subagents, and permissions | ⭐⭐⭐⭐ | Built on LangGraph; may be too heavy for simple agents. |
Microsoft family microsoft/agent-framework | New Python/.NET Microsoft agent projects | ⭐⭐⭐⭐ | Official migration path from AutoGen/Semantic Kernel. |
Microsoft family microsoft/autogen | Maintaining existing group-chat or debate projects | ⭐⭐⭐⭐ | Community-maintained maintenance mode; new projects use Agent Framework instead — skip the old 0.2 tutorials. |
Educational openai/swarm | Want to read a small source to understand handoffs | ⭐⭐⭐⭐ | Frozen for education; superseded by the Agents SDK, not for new production projects. |
Ecosystems langchain-ai/langchain | Want high-level model/retrieval/tool blocks | ⭐⭐⭐⭐ | Complex orchestration can drop down to LangGraph. |
Ecosystems run-llama/llama_index | Document-heavy retrieval workflows | ⭐⭐⭐ | Strong on data and retrieval, not every orchestration scenario. |
Counter-evidence Cognition — Don't Build Multi-Agents | Before adding a second agent | ⭐⭐⭐⭐ | Context fragmentation: once details scatter across agents, overall judgement can get worse. |
🛠 Hands-on practice (upstream)
Full exercises & starter codeSummaries from the upstream curriculum; full code, cost, and latency estimates live upstream.
- Exercises 1–2: fixed routes and conditional branches — draw the multi-step work as nodes/edges, starting from the LangGraph quickstart.
- Exercise 3: the supervisor pattern — a small two-role flow (researcher + writer); watch the handoffs and state.
- Exercise 4: CodeAct and sandboxes — compare generated code against tool calls, executed in isolation.
- Exercise 5: a type-safe agent — Pydantic AI returns answer/confidence/sources, and you watch non-compliant data get rejected.
- Recommended mini-project: a research-summary flow with a human gate — persist state, resume from a checkpoint, and stop at HITL until you approve the final publish step.
✅ Self-check
- I can tell an agent loop, an agent framework, a workflow graph, and multi-agent apart as different things.
- I start with the simplest option and add agents only on measurable evidence.
- I can explain what state, checkpoints, handoffs, and HITL each store or control.
- I ran the five exercises' offline tests and completed at least one Ollama Path A.
- I know CodeAct must run isolated and that type-safe output still needs content checks.
Adapted from awesome-agentic-ai-zh (MIT, by Wenyu Chiou) v2026.09.23; links checked 2026-08-27. Stars mark learning priority (⭐⭐⭐⭐⭐ = you will get stuck without it), not popularity. MIT License · Curriculum structure last updated 2026-10-03. Content is still being filled in; lessons marked “in progress” are not live yet.