Choose the right AI-agent framework before you build by translating requirements into architecture criteria, testing representative behavior, comparing tradeoffs, and defending a maintainable stack decision.
A practical, architecture-first course for choosing an AI-agent framework before committing to implementation. Learners translate system requirements into selection criteria, investigate the current framework landscape from official evidence, compare representative execution and integration models, design controlled proof-of-concept tests, and defend a maintainable stack decision. The course uses a refreshable comparison set rather than a frozen popularity list. At research time, representative evidence included graph-oriented libraries such as LangGraph; type- and schema-oriented frameworks such as PydanticAI; provider-associated kits such as Google ADK and the OpenAI Agents SDK; Microsoft Agent Framework; multi-agent or workflow-oriented offerings such as CrewAI, Strands Agents, and LlamaIndex workflows; and direct model SDK or custom application logic as an explicit baseline. Current documentation shows materially different approaches to persistence, handoffs, durable execution, context, tracing, model abstraction, and hosted services, so the course teaches learners to verify version-specific behavior rather than infer parity from similar terminology. ([openai.github.io](https://openai.github.io/openai-agents-python/agents/?utm_source=openai)) Mastery threshold: 80%.
What you will be able to do
- Produce a versioned requirements brief that translates an AI-agent use case into assessable framework-selection criteria covering execution control, state, tools, models, structured data, human review, evaluation, deployment, security, runtime, team capability, and maintenance.
- Classify current agent frameworks, SDKs, orchestration libraries, model-provider kits, hosted platforms, and custom/direct-SDK approaches without treating unlike categories as interchangeable.
- Compare representative frameworks by execution model, state boundaries, persistence, concurrency, routing, handoffs, inspectability, stop conditions, and extension points using current official evidence.
- Evaluate tool contracts, context handling, memory, structured outputs, retrieval, dependency injection, middleware, MCP support, provider portability, and integration lock-in through representative behavior tests.
- Design and execute framework proof-of-concept tests for reliability, evaluation, observability, human control, security, failure recovery, deployment fit, performance, cost, and maintainability.
- Separate measured evidence, documented capability, engineering inference, and stakeholder preference in a defensible selection scorecard.
- Recommend a framework, composite stack, direct model SDK, or custom application approach and defend the decision through an architecture decision record and implementation or migration plan.
- Assemble an AI Agent Framework Decision Portfolio demonstrating at least 80% mastery across requirements analysis, comparative evaluation, proof-of-concept evidence, and decision defense.
The AI Agentic School Decision Lab
Don’t just compare framework feature lists — make the stack decision and defend it. Each module ends with a realistic framework-selection brief. You create a concrete requirements matrix, architecture comparison, proof-of-concept plan/result, operational tradeoff review, or framework recommendation, explain your rationale, receive rubric-based coaching and a simulated stakeholder challenge, then revise until you demonstrate 80% mastery. Your best module submissions become your personal AI Agent Framework Decision Portfolio.
The curriculum is source-traceable and human-reviewed before publication. AI is used where it adds learner value—as a research tool and interactive practice/evaluation partner—not as a claimed framework vendor, benchmark authority, employer, or accreditation body.