AI Orchestrators
This subject teaches engineers how to build, coordinate, and operate multi-agent LLM systems rather than single-shot prompts. The core stack includes OpenAI function/tool calling, structured outputs, LangGraph and LangChain for stateful control flow, AutoGen and CrewAI for role-based agent teams, and Model Context Protocol for exposing tools. Students implement common orchestration topologies — supervisor-worker, peer debate, group chat, and reflection loops — and learn where each topology fails around token cost, message fan-out, and context overflow. The second half focuses on production behavior: memory design with vector stores and short-term scratchpads, observability with Langfuse/OpenTelemetry traces, evaluation with trajectory and tool-call assertions, and human-in-the-loop approval gates. The program is hands-on with Python and current agent SDKs; students finish by packaging a resumable multi-agent system with guardrails, retries, and evals.
What You'll Learn
Curriculum
Builds the runtime loop that drives every orchestrator: model call, tool call, result, repeat. Students implement function schemas, stop-reason handling, and error recovery against OpenAI-compatible APIs.
- The Single-Agent Control Loop: Model Call, Tool Call, Result, Repeat 3.0h
- Function Schemas and Stop-Reason Handling for Tool Dispatch 3.0h
- Error Recovery, Retries, and Guardrails for Tool-Calling Loops 3.5h
Moves from single loops to durable, branching agent workflows using LangGraph's state machine primitives. Students build resumable graphs with conditional routing, cycles, and checkpointing for a research-and-summarize agent.
- Lesson 1: Building Your First Stateful Graph: Nodes, Typed State, and Conditional Routing 2.5h
- Lesson 2: Cycles, Critique, and Human-in-the-Loop Interrupts 3.0h
- Lesson 3: Durable Checkpointing and Resumable Execution 3.0h
Implements supervisor-worker, peer debate, and group chat topologies in AutoGen and CrewAI. Students measure task ownership, token overhead, and failure modes for each pattern.
- Supervisor-Worker Topology in AutoGen and CrewAI 2.5h
- Peer Debate and Consensus Patterns 2.5h
- Group Chat Topologies and Cross-Framework Evaluation 3.0h
Covers Model Context Protocol servers, pgvector-backed semantic memory, and permissioned tool execution. Students expose an internal API through MCP and add short-term scratchpad plus long-term retrieval to a support agent.
- Exposing Internal APIs as MCP Tools with Scoped Permissions 3.0h
- Short-Term Scratchpad and Context Assembly for Support Agents 2.5h
- pgvector-Backed Long-Term Semantic Memory with Namespaced Retrieval 3.0h
Instruments multi-agent runs with Langfuse and OpenTelemetry traces, then builds assertion-based and agent-as-judge eval suites. Students add human-in-the-loop approvals, retries, rate limiting, and cost tracking before shipping an orchestrated system.
- Instrumenting Multi-Agent Systems with OpenTelemetry and Langfuse 3.0h
- Building Evaluation Suites: Assertions and Agent-as-Judge 3.0h
- Production Controls: Human-in-the-Loop, Retries, Rate Limiting, and Cost Management 2.5h
Career Relevance
Graduates can build agent workflows in LangGraph and AutoGen with explicit state, tool schemas, and supervisor controls, rather than chaining brittle prompts. They can expose internal services through Model Context Protocol and set up Langfuse/OpenTelemetry traces plus assertion-based evals for regressions. The work matches AI engineer, platform/backend engineer on agent infrastructure, and ML platform roles where teams deploy multi-agent systems behind human approval gates.
Recommended Background
Working knowledge of Python, REST/JSON APIs, and calling LLM chat or completion endpoints; familiarity with basic prompt design. Prior experience with LangChain, AutoGen, or vector databases is helpful but not required.
Program Details
| Award | Certificate |
| Duration | 6 months |
| Level | Intermediate |
| Courses | 5 |
| Language | English (UI in 4 languages) |
Caden Academy is a private, fully online professional training provider operated by Replatform (a company registered in Hong Kong). We are not a university, college, or accredited institution.
We do not hold any educational license, government approval, or accreditation from any authority in any country — including the United States, Hong Kong, China, or the European Union.
The certificates and diplomas we issue are issued solely by us, as a private company. They are not recognized by governments, universities, or official qualification frameworks. They cannot be used for:
- Academic credit transfers to any college or university
- Immigration or visa applications
- Government employment or civil service purposes
- Professional licensing requirements in any jurisdiction
Our programs are designed for personal and professional development only. Whether an employer chooses to recognize our certificates is entirely at the employer's discretion, and we make no guarantees about career outcomes, salary changes, or employment prospects.