BERDAU EMPIREAI for creators

Track 02 — Systems

Multi-agent AI workflows

An agent is not a smarter chatbot. It is a worker with a brief, tools, and a stopping condition. Beginners skip the third part and build a liability.

Intermediate 4 modules 4 weeks 4–5 hrs / week

A chat turn proposes something and waits. An agent acts, looks at the result, and decides what to do next — without you in the loop for every step. That difference is the whole value, and the whole risk.

This track is deliberately narrow: you build one agent, properly. Not a swarm, not a framework tour. One worker with a written brief, three tools, a budget, and a test suite that tells you when it's lying to you.

What you walk away with

  • A written brief for one agent, with stopping conditions that are actually checkable
  • A tool set limited by least privilege, with irreversible steps behind confirmation
  • A twelve-case eval with a ship threshold and a rollback trigger
  • A log that lets you reconstruct why it did what it did
Reality check

Most creators do not need agents yet. If you haven't run the workflow manually twenty times, you will automate a process you don't understand — and get faster at the wrong thing. Do the manual version until it's boring. Then automate the boring part.

The four modules

  1. What an agent actually is

    Brief, tools, stopping condition. We look at three real failures — a loop that never ends, an action nobody approved, and a confident report built on an empty source — and trace each back to the missing part.

    You produce: a one-page map of a workflow you already run by hand.

  2. Writing the brief

    Role, inputs, procedure, stopping conditions, output contract, blast radius. The hard one is stopping conditions: "when unsure" is not a condition, it's an absence of one.

    You produce: a brief where every step names where its inputs come from.

  3. Tools and permissions

    Least privilege, argument validation, and the rule that anything irreversible needs a human yes. We also wire a log you can actually read a week later.

    You produce: a tool list where each one has a stated worst-case action.

  4. The eval that decides whether it ships

    Twelve cases — four typical, four edge, four adversarial — with pass criteria a second person could apply. Two of the twelve must be cases where the correct answer is "I can't do that."

    You produce: an eval sheet, a ship threshold, and a rollback trigger.

What this track is not

It is not a framework recommendation. The brief survives a framework change; the framework doesn't. It also isn't a shortcut to content volume — an agent that posts eight times a day will happily post eight pieces of nothing.

Where to go next

Read the agent brief note first if you want the template before the theory. If your visual output is still inconsistent, go back to content & visual production — the system can't fix a look that was never defined.