Mariete Builder Specialist
Foundation taught you to brief, review, and schedule a single Builder run. Specialist is what you reach for when one agent is no longer the right unit of work. Multi-agent decomposition, SDK-built sub-agents with their own tool boundaries, human-in-the-loop approval gates, versioned rollouts with regression detection, and an incident practice that survives a production failure. This course assumes Foundation; it skips the basics and gets directly to the engineering decisions that separate a hobby workspace from a system you can defend in front of a CTO, an auditor, and a regulator.
What you'll be able to do.
- Multi-agent decomposition
- Custom SDK agents
- Approval gates and human-in-the-loop
- Production governance
- Failure modes and incident response
5 modules. ~110 minutes total.
- 01Multi-agent decomposition~22 min
A single Builder run is the right unit of work for a job that fits in one head. Once a job needs two heads — a researcher and a writer, a tr…
- 02Custom SDK agents~22 min
The 50+ pre-built recipes cover most jobs. The Builder SDK is for the jobs they don't. This module is about when to drop into the SDK, how t…
- 03Approval gates and human-in-the-loop~22 min
Some agent decisions must not happen without a human. Sending external email to a regulator. Posting to the company's official social accoun…
- 04Production governance~22 min
A Builder workspace with three schedules and one operator is easy to govern. A workspace with sixty schedules across eight teams, three mode…
- 05Failure modes and incident response~22 min
Production multi-agent systems fail. Not "might fail" — will fail. The team that builds Builder pipelines as if failure is impossible has no…
One sample question. Click to reveal the answer.
A teammate proposes splitting a 5-step lead-qualification job into 5 separate agents because "more agents are more flexible." What is the most accurate response, given the decomposition criteria?
- Approve — granular agents are always more reviewable than monolithic ones
- Reject the split unless at least two of the five decomposition criteria are met (distinct phases, isolation, different approvers, different cadences, divergent token budgets)
- Approve only if the team also creates a "skeptic agent" to review each step
- Reject — multi-agent flows are only appropriate for SDK-built agents, never for forks
20 questions like this. Builder Specialist pass-mark is 75%.