Build the smallest agent that reliably creates value
A platform-neutral foundation for deciding, designing, constraining, evaluating, and operating AI agents. You'll learn how to tell an agent apart from a workflow, reject one when a simpler design wins, and carry a single enterprise use case from problem to a controlled, costed, portable design.
The one line to remember: Workflows decide the path. Agents decide part of the path. Tools do the work. Policies decide what is allowed.
What outcome matters, and why is an agent the smallest sufficient solution?
How should the work change, and what evidence does the system need?
What may the agent know, decide, and call — and how will we prove it works?
What must be prevented, approved, detected, reversed, or escalated?
Is it reliable, observable, affordable, supportable, and portable?
- 1Deterministic automation — Stable rules, calculations, queries, scheduled jobs, and predictable process paths.
- 2Single model call — Bounded generation, classification, extraction, or summarization with no adaptive loop.
- 3Retrieval assistant / RAG — Grounded question answering or synthesis when the path is fixed and actions are read-only.
- 4LLM workflow — Probabilistic work helps, but the control path can stay coded and testable.
- 5Bounded single agent — Judgment, exceptions, unstructured evidence, or paths that can't be reliably predetermined.
- 6Multi-agent system — Only when specialization, context isolation, or cross-boundary ownership defeats a single agent.
Test solutions bottom-up. Climb a rung only when the simpler one provably can't meet the outcome, exception, or quality requirement at a better risk-adjusted cost.
Cost per accepted outcome — the full cost (model, tools, review, rework, failures, controls, assurance, and support) of one outcome that clears quality, latency, safety, and control thresholds.
- Knowledge & cumulative retrieval checks15%
- Module portfolio artifacts35%
- Evaluation & security practical20%
- Capstone simulation & design defense30%
- 80% overall
- No competency below 70%
- Mandatory pass on security-critical controls
~16 hours · 80% stable platform-neutral core, 20% updateable platform lenses.
Brought to you by KokoAI Academy — we build production AI systems like the ones this course is about. It's free, and platform-neutral: the doctrine travels to any major platform, and no vendor account is needed.