Live LabsBuildersSystem design2 hours

Pick the right architecture: single agent, workflow or multi-agent

Most agent projects fail at the first decision: how many agents, and who decides what runs next. We build the same task three ways and measure what each costs in tokens, latency and failure.

You leave with

A decision table and a measured comparison of two designs for one real task.

How the time is spent

  1. I do30 minParam builds it live

    I run one support-triage task as a single agent, as a fixed workflow and as a supervisor with two workers, and show the traces side by side.

  2. We do48 minTogether, step by step

    We instrument all three, count tokens and latency, and break each one on purpose to see how it fails.

  3. You do42 minOn your own code

    You pick a design for a task from your own work, run it, and defend the choice with your numbers.

What you will be able to do

Choose the shape first

Tell when a workflow beats an agent, and when a second agent earns its cost.

Measure, do not guess

Compare designs on tokens, latency and failure rate from real runs.

Write the decision down

Leave with a one-page decision record your team can review.

Starts from

Before you come

For
Engineers about to build their first or second agent at work.
You need
Comfortable with TypeScript or Python. An LLM API key.
Stack
TypeScript, Mastra, LangGraph
Track
Hands-on agentic AI for engineers who ship.

More builders labs