AI Engineering: Agents

AI Engineering: Agents is a free two-hour course from the Weights & Biases AI Academy, produced in collaboration with OpenAI and taught by Ilan Bigio, a developer-experience engineer at OpenAI, and Anish Shah, an AI engineer at Weights & Biases. It is organised as six modules, each adding one capability to the same running codebase rather than starting over: deterministic LLM workflows built from chaining and structured outputs, where reliability comes from constraining the model; a single-agent system that decides autonomously when to call a tool; context, memory and retrieval, giving the agent state across turns; multi-agent collaboration through orchestrator-worker patterns and structured hand-offs; evaluation and benchmarking against accuracy, latency and cost; and finally MCP, so the agent can reach tools it did not ship with. The code lives in the public wandb/agents-course repository as one Python file per module — _1_workflow.py, _2_agent.py, _3_memory_retrieval.py, _4_multi_agents.py, _5_evals.py with a lighter _5_simple_evals.py alongside it, and _6_mcp.py — written against the OpenAI API with W&B Weave for tracing and evaluation. Setup is documented for uv, pip, pyenv and conda, and needs Python 3.11 or newer, an OpenAI API key, a Weights & Biases API key, and npx for the MCP module. The framing throughout is production reliability rather than demo-building, which is why benchmarking gets a module of its own instead of a closing mention.

Topic

Agentic

Level

Intermediate

Format

Course

Cost

Free

Duration

~2 hours, self-paced (6 modules)

Provider

Weights & Biases

Hands-on

Yes — code/exercises

Certificate

None

• ✓Build deterministic LLM workflows with chaining and structured outputs first
• ✓Implement a minimal single agent that decides when to call its tools
• ✓Add memory storage and retrieval so an agent keeps state across turns
• ✓Coordinate multiple agents with orchestrator-worker patterns and structured hand-offs
• ✓Benchmark agents on accuracy, latency and cost in a reproducible way
• ✓Trace every agent call with W&B Weave to see where a run went wrong
• ✓Attach MCP servers so an agent reaches tools outside its own codebase
• ✓Set up the project with uv, pip, pyenv or conda and the required API keys

• •Co-produced with OpenAI and co-taught by an OpenAI developer-experience engineer
• •One codebase grown across six modules, so you see why each layer was added rather than meeting it fully formed
• •All code is public in the wandb/agents-course repository, one Python file per module
• •Starts with deterministic workflows before agents — the honest ordering, since most tasks do not need autonomy
• •Evaluation and benchmarking gets a full module against accuracy, latency and cost, not a closing slide
• •Two hours and free, with no certificate advertised — this is skills, not a credential

This page was written from 2 sources, 1 on domains other than wandb.ai.

1. 1.wandb.ai — agentsvendor
2. 2.github.com — agents course