FDE COMMONS/by AIR

Agentic AI System Design / Sources

Primary sources

Anything an engineer would act on is cited to official documentation, not to a blog post about the documentation. Framework behavior, protocol semantics, and provider limits move faster than any course, so check the current version of a page before you build against it.

Where a vendor blog and the vendor's own docs disagree, the docs win. Where the docs are silent, the course says so rather than guessing.

Course version 0.1.0 · sources last reviewed 2026-08-26

Purpose: Keep implementation-sensitive course claims tied to official documentation, standards, and original engineering publications.
Last reviewed: 2026-08-26

SDKs, protocols, model behavior, pricing, and product features change. Re-check implementation details before publishing a dated code lab. The architectural principles in the lessons should remain useful even when adapters change.

How to use this page

  • Use the lesson as the conceptual source of truth for the course.
  • Use these links to verify current API names and implementation behavior.
  • Prefer primary sources over blog summaries.
  • Record a new review date when changing framework-specific code.
  • Do not infer production guarantees from a feature name. Test the behavior.

Agent architecture and control patterns

SourceCourse useModules
Anthropic: Building Effective AgentsWorkflows versus agents; simple composable patterns; routing, parallelization, orchestrator-worker, evaluator-optimizer01, 02, 06, 09, 15
OpenAI Agents SDKAgents, tools, handoffs, guardrails, sessions, tracing, and implementation adapter02, 03, 05, 08, 09, 12
OpenAI Agents SDK: Running AgentsAgent loop behavior, turns, execution, and integration boundaries02, 07
LangGraph: Workflows and AgentsExplicit graph patterns and agent/workflow distinctions02, 06
LangGraph OverviewGraph runtime, durable execution, human control, and memory adapter06, 07, 08

Structured outputs and tool use

SourceCourse useModules
OpenAI: Function CallingTyped tool calling and application execution02, 03
OpenAI: Structured OutputsSchema-constrained decisions and artifacts02, 03, 11
Anthropic: Tool UseTool schemas and provider-specific adapter behavior02, 03
OpenAI Agents SDK: ToolsTool wrappers, hosted tools, function tools, and adapter examples03
OpenAI Agents SDK: GuardrailsInput/output/tool guardrail adapter behavior02, 10

Context engineering, retrieval, and memory

SourceCourse useModules
Anthropic: Effective Context Engineering for AI AgentsContext selection, compaction, tool-result handling, and long-horizon context04, 05, 13
LangChain: Short-Term MemoryThread-scoped state and context management adapter05
LangChain: Long-Term MemoryCross-thread memory concepts and implementation adapter05
LangChain: Memory OverviewMemory categories and framework vocabulary05
OpenAI Agents SDK: SessionsSession history adapter05

Durable and long-running execution

SourceCourse useModules
Temporal: Understanding Durable ExecutionEvent history, recovery, replay, and durable workflow model07
Temporal: WorkflowsDeterministic workflow code and lifecycle07
Temporal: ActivitiesExternal work, retries, idempotency, and activity boundaries03, 07
Temporal: Activity Definition and Retry PolicyTimeouts, retries, heartbeats, and errors07, 13
Temporal: Event HistoryEvent-history model and replay explanation07
Temporal: Worker VersioningSafe deployment of long-running workflows07, 14
LangGraph: PersistenceCheckpoints, threads, and runtime persistence adapter05, 07
LangGraph: InterruptsPause/resume and external input adapter07, 08

Human-in-the-loop

SourceCourse useModules
OpenAI Agents SDK: Human in the LoopApproval and resumption adapter behavior08
LangChain: Human-in-the-LoopMiddleware and interrupt patterns for tool review08
LangGraph: InterruptsDurable pause and resume07, 08

Multi-agent and interoperability

SourceCourse useModules
OpenAI Agents SDK: HandoffsHandoffs versus agents-as-tools09
Model Context Protocol: IntroductionMCP purpose and ecosystem vocabulary09
Model Context Protocol: ArchitectureHost/client/server boundaries and capability negotiation09, 10
Model Context Protocol: Security Best PracticesAuthorization, confused-deputy, token, and trust-boundary guidance09, 10
Agent2Agent ProtocolCurrent A2A specification and reference implementations09
Google: Announcing A2AOriginal protocol rationale and cross-vendor task model09

Security, isolation, and governance

SourceCourse useModules
OWASP Top 10 for Agentic Applications 2026Agent-specific threat categories and mitigation review03, 05, 09, 10, 14
OWASP Top 10 for LLM ApplicationsLLM application risks04, 10
OWASP: Prompt InjectionDirect and indirect injection04, 10
NIST AI Risk Management FrameworkGovern, map, measure, and manage risk00, 10, 14, 15
NIST AI RMF: Generative AI ProfileGenerative-AI risk considerations and controls10, 14
Docker Engine SecurityContainer isolation limits and hardening context10
Docker AI Sandboxes: IsolationSandbox isolation for agent-executed development tasks10

Evaluation and testing

SourceCourse useModules
OpenAI: Evaluate Agent WorkflowsAgent evaluation concepts and current platform adapter11
OpenAI: Trace GradingTrajectory and trace-level grading11, 12
OpenAI: Evaluation Best PracticesDataset and grader design11
Anthropic: Demystifying Evals for AI AgentsAgent-eval methodology and failure analysis11
OpenAI Agents SDK: TracingTrace capture and processor adapter11, 12

Observability and reliability

SourceCourse useModules
OpenTelemetry: Generative AI Semantic ConventionsCurrent semantic-convention vocabulary; verify stability status before hard-coding12
OpenTelemetry: Observing Generative AI SystemsGenAI observability rationale and ecosystem context12
Google SRE Book: Service Level ObjectivesSLIs, SLOs, and error budgets12
Google SRE Book: Handling OverloadAdmission control, load shedding, and overload behavior13

Latency, cost, and scale

SourceCourse useModules
OpenAI: Latency OptimizationCurrent provider-specific latency techniques13
OpenAI: Prompt CachingPrompt-cache behavior and constraints13
OpenAI: Batch APIOffline/batch processing adapter13
Anthropic: Prompt CachingProvider-specific caching adapter13

Deployment and production operations

SourceCourse useModules
OpenAI: Production Best PracticesCurrent API production guidance13, 14
Temporal: Worker VersioningLong-running workflow deployment07, 14
NIST AI Risk Management FrameworkGovernance and operational lifecycle10, 14

Role competency anchors

These sources should inform positioning and interview outcomes, not become a keyword checklist.

SourceWhy it matters
OpenAI: Applied AI Engineer, EnterpriseIllustrates the breadth expected across models, agents, retrieval, tools, reliability, observability, latency, cost, safety, security, and governance
OpenAI: Forward Deployed EngineerIllustrates end-to-end design, deployment, and customer-environment ownership
Anthropic CareersUse current role pages when publishing role-specific guidance
Google CareersUse the current job description when tailoring interview modules to a specific opening

Review checklist for framework-specific labs

Before publishing or updating a lab, verify:

  1. The linked page still exists.
  2. The feature is stable enough for the lesson.
  3. The code uses current package and API names.
  4. Authentication and data-retention behavior are documented.
  5. Tool approval and tracing behavior are tested, not assumed.
  6. Durable execution semantics are demonstrated with a crash test.
  7. Protocol security guidance is reflected in the example.
  8. The lab preserves the architecture-first explanation.
  9. Provider pricing or limits are dated and kept out of timeless lesson claims.
  10. The page’s last_reviewed field is updated.