Skip to main content

Study Guide

Purpose of this guide​

The SSA Core Track teaches you to design, build, operate, and scale semantic AI systems. But no curriculum can cover everything, and the field evolves faster than any course can be updated.

This guide helps you deepen your knowledge beyond the curriculum. For each topic area aligned with the 8 core modules, you will find the key concepts to study, the ideas that matter most, practice exercises to sharpen your skills, and connections to adjacent fields that enrich your perspective.

This is not a reading list of specific titles. It is a map of ideas. Seek out the sources that explain these ideas in ways that work for you -- textbooks, articles, courses, talks, documentation, or conversations with practitioners.

Topic Area 1: Knowledge Representation and Ontologies​

Aligned with Module 2 -- Semantic Modeling and Ontologies.

Key concepts to study​

  • Taxonomies, thesauri, and ontologies. Understand the spectrum from simple classification systems (taxonomies) through controlled vocabularies with relationships (thesauri) to formal representations with axioms and inference rules (ontologies). Know when each level of formality is appropriate.
  • Entity-relationship modeling. Study the fundamentals of identifying entities, attributes, and relationships in a domain. This is a foundational skill from database design that translates directly to ontology construction.
  • Description logics and formal semantics. Learn the basics of how formal ontology languages represent classes, properties, and constraints. You do not need to become a logician, but understanding what formal semantics offers helps you decide how much formality your ontology needs.
  • Knowledge graphs. Study how organizations use graph structures to represent interconnected knowledge. Understand nodes, edges, properties, and traversal patterns.
  • Controlled vocabularies and naming conventions. Learn how standardized terminology reduces ambiguity in large organizations. Study how healthcare, law, and finance maintain controlled vocabularies.

Ideas that matter most​

The single most important idea is that meaning must be made explicit. AI systems cannot infer what you left unstated. The second most important idea is that all models are simplifications, and the art is knowing what to leave out. The third is that ontologies are living documents that must evolve with the domain.

Practice exercises​

  1. Pick a domain you know well (a hobby, your workplace, a sport) and build a complete ontology from scratch. Include at least 15 entities, their relationships, lifecycle states, and 5 invariants. Then ask someone else who knows the domain to review it and note every place where your model surprised them.
  2. Take two competing ontologies for the same domain (for example, two different ways to classify customer support requests) and write a comparison that explains what each one gets right, what it misses, and what would happen if an AI system used each one.
  3. Version your ontology three times, simulating a domain change each time (a new product category, a regulatory change, a merger with another company). Document the impact of each change on downstream contracts.

Connections to adjacent fields​

  • Library science and information architecture have decades of experience organizing knowledge for retrieval.
  • Cognitive science studies how humans categorize and reason about concepts, which informs how to build ontologies that match human mental models.
  • Database design provides mature techniques for entity-relationship modeling and normalization that apply directly to ontology construction.

Topic Area 2: Multi-Agent Systems​

Aligned with Module 3 -- Agent Architecture.

Key concepts to study​

  • Agent theory. Study the foundational concepts of autonomous agents: perception, reasoning, action, and communication. Understand the difference between reactive agents and deliberative agents.
  • Coordination mechanisms. Learn how multiple agents can coordinate their behavior through shared state, message passing, contract-based negotiation, and organizational structures.
  • Role-based design. Study how responsibilities are allocated in multi-agent systems. Understand the tradeoffs between specialization (agents that do one thing well) and generalization (agents that handle many tasks).
  • Failure modes in distributed systems. Learn about partial failures, Byzantine faults, consensus problems, and the strategies that distributed systems use to maintain correctness. Many of these concepts apply directly to multi-agent AI systems.
  • Organizational theory. Study how human organizations structure teams, delegate authority, and coordinate work. The analogy between agent architectures and organizational design is deep and productive.

Ideas that matter most​

The most important idea is that the architecture of the team determines the quality of the outcome. A brilliant agent in a poorly designed system will produce poor results. The second key idea is that every handoff is a potential failure point, so handoff design deserves as much attention as agent design. The third is that you should design for failure from the start, not as an afterthought.

Practice exercises​

  1. Design three different agent architectures for the same problem (for example, a travel booking assistant) using supervisor, pipeline, and swarm topologies. Write a decision memo explaining which one you would choose and why.
  2. Take an existing multi-agent design and deliberately break it by removing one agent. Document what happens to the system and how you would redesign it to be resilient to this failure.
  3. Conduct a failure mode analysis on a multi-agent system you have designed. Identify at least 10 failure scenarios and design detection and recovery mechanisms for each.

Connections to adjacent fields​

  • Microservices architecture faces similar challenges around service boundaries, inter-service communication, and failure handling.
  • Team management and organizational design provide models for how to structure groups of specialized workers.
  • Game theory offers frameworks for understanding how agents with different objectives interact.

Topic Area 3: Prompt Engineering and Context Design​

Aligned with Module 5 -- Context Engineering and Structural Prompting.

Key concepts to study​

  • Instruction following and instruction hierarchy. Understand how language models process instructions, how instruction order affects behavior, and how to structure instructions for maximum clarity.
  • Few-shot learning and in-context learning. Study how examples in the prompt influence model behavior. Learn when few-shot examples help, when they hurt, and how to select them.
  • Chain-of-thought and reasoning strategies. Learn how prompting techniques that encourage step-by-step reasoning improve output quality for complex tasks.
  • Output formatting and schema enforcement. Study techniques for constraining model output to specific formats, including JSON schemas, XML structures, and structured text.
  • Context window management. Understand how the finite context window constrains what information the agent can access, and how to prioritize what goes in and what stays out.
  • Prompt versioning and lifecycle management. Learn how to treat prompts as versioned artifacts with testing, review, and deployment processes.

Ideas that matter most​

The most important idea is that context is the entire design surface for AI behavior. What the model sees determines what it does. The second key idea is that structure beats cleverness -- a well-organized mediocre prompt outperforms a brilliant but chaotic one. The third is that prompts drift, and what works today may not work after the next model update.

Practice exercises​

  1. Take a single complex prompt and decompose it into modular components: identity, policies, task instructions, output schema, and refusal rules. Measure whether the modular version produces more consistent results.
  2. Create a consistency test suite for a prompt by running the same 20 inputs through it 5 times each and measuring output variance. Then apply structural improvements and re-test.
  3. Design a context package for an agent that must operate within a strict token budget (for example, 4,000 tokens for the system prompt). Justify every token allocation decision.

Connections to adjacent fields​

  • Technical writing provides principles for clear, unambiguous instruction design.
  • User interface design offers frameworks for information hierarchy and progressive disclosure that apply to context architecture.
  • Curriculum design shows how to sequence information for optimal comprehension, which is relevant to how you structure an agent's instructions.

Topic Area 4: Software Quality and Testing​

Aligned with Module 6 -- Evaluation, Quality and Observability.

Key concepts to study​

  • Testing theory. Study the fundamentals of test design: equivalence partitioning, boundary value analysis, decision tables, and state transition testing. These classic techniques apply to AI system evaluation with adaptation.
  • Metrics design. Learn how to define metrics that are valid (they measure what they claim to measure), reliable (they produce consistent results), and actionable (they guide specific decisions).
  • Observability engineering. Study the three pillars of observability (logs, metrics, traces) and how they combine to provide understanding of system behavior. Learn how observability differs from monitoring.
  • Statistical quality control. Understand concepts like process capability, control charts, and statistical significance. These help you distinguish real quality changes from random variation.
  • Regression testing strategies. Learn how to build test suites that catch degradation efficiently without becoming so large that they slow down development.

Ideas that matter most​

The most important idea is that if you cannot measure it, you cannot manage it. The second is that AI quality is probabilistic, not binary -- the system does not simply "work" or "not work," it works at a certain quality level that can drift. The third is that the eval suite is a living artifact that must grow and evolve with the system.

Practice exercises​

  1. Design an eval suite for a system you use daily (for example, a search engine or a recommendation system). Define 30 test cases covering happy path, edge cases, and adversarial scenarios. Write rubrics for each.
  2. Build a metric contract that maps a business KPI to quality indicators to technical signals for a specific AI system. Define thresholds, measurement frequency, and ownership.
  3. Simulate a semantic regression by changing one parameter in a system (for example, swapping a prompt section) and then using your eval suite to detect the change. Document the detection process.

Connections to adjacent fields​

  • Six Sigma and lean manufacturing provide mature frameworks for quality measurement and continuous improvement.
  • Psychometrics offers techniques for designing assessments that are valid and reliable, directly applicable to rubric design.
  • Epidemiology provides models for detecting changes in population-level outcomes that parallel how we detect quality drift in AI systems.

Topic Area 5: Information Security​

Aligned with Module 7 -- Security, Governance and Ethics.

Key concepts to study​

  • Threat modeling methodologies. Study STRIDE, PASTA, LINDDUN, and other structured approaches to identifying threats. Understand how to adapt these for AI-specific attack surfaces.
  • Prompt injection and adversarial attacks. Learn the taxonomy of attacks specific to language model systems: direct prompt injection, indirect prompt injection, data poisoning, and model manipulation.
  • Defense in depth. Study the principle of layered security controls, where no single control is expected to stop all attacks. Understand how input validation, output filtering, access controls, and monitoring work together.
  • Access control models. Learn about role-based access control (RBAC), attribute-based access control (ABAC), and the principle of least privilege. Understand how these apply to AI systems where agents act on behalf of users.
  • Audit and compliance. Study how audit trails are designed to satisfy regulatory requirements. Understand the difference between logging for debugging and logging for compliance.

Ideas that matter most​

The most important idea is that AI systems have attack surfaces that traditional software does not, because the instruction layer is exposed to user input. The second is that security is an architectural concern, not a feature to add later. The third is that perfect security is impossible, so the goal is to make attacks detectable and recoverable.

Practice exercises​

  1. Conduct a STRIDE analysis on an AI system you have designed. For each threat category, identify at least two specific attack scenarios and design mitigations.
  2. Write 20 adversarial prompts designed to make an agent violate its policies. Then design guardrails that would catch each one.
  3. Design an access control model for a multi-agent system with three user roles (customer, support agent, administrator) and five agent types. Define exactly what each user role can trigger each agent to do.

Connections to adjacent fields​

  • Application security (AppSec) provides mature practices for secure development lifecycle, code review, and vulnerability management.
  • Privacy engineering offers frameworks for data minimization, purpose limitation, and consent management.
  • Red teaming provides methodologies for adversarial testing that can be adapted for AI systems.

Topic Area 6: Product Management​

Aligned with Module 8 -- Product, Operations and Scale.

Key concepts to study​

  • Product discovery and validation. Study how to identify user needs, validate assumptions, and iterate toward product-market fit. Understand how these practices adapt when the product includes AI components.
  • Roadmap design. Learn how to create roadmaps that align technical capabilities with business outcomes and user needs. Understand the difference between feature-driven and outcome-driven roadmaps.
  • Stakeholder management. Study techniques for communicating with diverse stakeholders (engineers, executives, users, regulators) who have different priorities and different levels of technical understanding.
  • Value proposition design. Learn how to articulate the specific value an AI system creates, measured in terms the business understands (cost savings, revenue growth, risk reduction, time savings).
  • Build versus buy versus integrate decisions. Understand the frameworks for deciding when to build custom AI systems, when to use off-the-shelf solutions, and when to integrate multiple components.

Ideas that matter most​

The most important idea is that an AI system is not a product until it delivers measurable value to someone. Architecture without product strategy is an academic exercise. The second key idea is that AI products have unique lifecycle characteristics (model drift, evolving capabilities, probabilistic behavior) that product management must account for. The third is that managing expectations is as important as managing technology.

Practice exercises​

  1. Write a one-page value proposition for an AI system, explaining what it does, who it serves, and how you would measure its success. Have someone without technical background read it and tell you what they understood.
  2. Create a two-quarter product roadmap for an AI system, linking each planned capability to a specific business outcome and a measurable success metric.
  3. Design a stakeholder communication plan for an AI system launch, with different messages for different audiences (engineering team, executive sponsors, end users, compliance team).

Connections to adjacent fields​

  • Jobs-to-be-done theory provides a framework for understanding what users actually need, which helps prioritize AI system capabilities.
  • Lean startup methodology offers practices for rapid experimentation and validated learning.
  • Change management provides models for helping organizations adopt new tools and practices, critical for AI system rollouts.

Topic Area 7: Site Reliability Engineering​

Aligned with Module 8 -- Product, Operations and Scale.

Key concepts to study​

  • SLOs, SLIs, and error budgets. Study how to define service level objectives, measure them with service level indicators, and use error budgets to balance reliability with development velocity.
  • Incident management. Learn structured approaches to incident detection, response, mitigation, and resolution. Study incident command systems and communication protocols.
  • Post-mortem culture. Understand how to conduct blameless post-mortems that produce genuine system improvements rather than blame assignments.
  • Capacity planning. Study how to predict and prepare for growth, including the cost implications of scaling AI systems (model inference costs, storage, retrieval infrastructure).
  • Runbook design. Learn how to write operational runbooks that enable on-call engineers to diagnose and mitigate common issues quickly and consistently.

Ideas that matter most​

The most important idea is that reliability is a feature, not an accident -- it must be designed, measured, and maintained. The second is that AI systems have unique operational characteristics (model drift, semantic regression, probabilistic behavior) that traditional SRE practices must be adapted to handle. The third is that the error budget is your best tool for balancing innovation with stability.

Practice exercises​

  1. Define SLOs for an AI system with at least three dimensions: availability, latency, and quality. For each SLO, define the SLI that measures it and the error budget that governs it.
  2. Write an incident runbook for a semantic regression scenario: the agent's responses are technically valid but quality has degraded. Include detection criteria, triage steps, mitigation options, and escalation paths.
  3. Conduct a capacity planning exercise for an AI system that needs to scale from 1,000 to 100,000 daily interactions. Estimate the cost implications and identify the bottlenecks.

Connections to adjacent fields​

  • DevOps practices provide foundations for continuous delivery, infrastructure as code, and automated testing.
  • Industrial operations management offers frameworks for managing complex systems that apply to AI operations.
  • Chaos engineering provides methodologies for proactively testing system resilience.

Topic Area 8: Ethics in AI​

Aligned with Module 7 -- Security, Governance and Ethics.

Key concepts to study​

  • Fairness and bias. Study the different definitions of fairness (demographic parity, equalized odds, individual fairness) and understand why they can conflict with each other. Learn how bias enters AI systems through data, design, and deployment decisions.
  • Transparency and explainability. Understand the spectrum from fully opaque systems to fully transparent ones. Study techniques for making AI decisions interpretable to different audiences.
  • Accountability and governance. Learn how to design accountability structures that clarify who is responsible when an AI system causes harm. Study governance models that include diverse stakeholders.
  • Value alignment. Understand the challenge of specifying human values in terms precise enough for a system to follow. Study how value trade-offs arise and how to make them explicit.
  • Human oversight and control. Learn about the spectrum of human involvement, from fully automated systems to human-in-the-loop designs. Understand when each level is appropriate.

Ideas that matter most​

The most important idea is that ethics is not a separate concern but an integral part of system design. Every architectural decision has ethical implications. The second is that there is no single correct ethical framework -- different situations require different ethical reasoning. The third is that making values explicit is better than leaving them implicit, even when the explicit values are imperfect.

Practice exercises​

  1. Take an AI system design and identify five ethical risks. For each risk, describe the harm that could occur, who would be affected, and how the system design could be modified to reduce the risk.
  2. Design a fairness audit process for an AI system that makes decisions affecting people (hiring, lending, content moderation). Define what fairness means in this context, how you would measure it, and what thresholds would trigger remediation.
  3. Create an ethical decision framework for a specific AI system that makes the values explicit, defines decision criteria for common ethical dilemmas, and establishes accountability for ethical outcomes.

Connections to adjacent fields​

  • Philosophy (ethics and epistemology) provides foundational frameworks for moral reasoning and knowledge.
  • Law and regulation defines the legal boundaries within which AI systems must operate.
  • Sociology and critical theory offer perspectives on power, inequality, and systemic effects that inform how we evaluate AI impact.

Study strategies​

How to approach this material​

  1. Breadth first, then depth. Start by getting a general understanding of each topic area. Then go deep in the areas most relevant to your current work or career goals.

  2. Theory and practice in balance. For every concept you study, try to apply it to a real or simulated system. Understanding without application fades quickly.

  3. Build a personal knowledge base. As you study, create your own notes, diagrams, and examples. The act of restating concepts in your own words deepens understanding.

  4. Seek multiple perspectives. No single source explains everything well. Read different authors, watch different speakers, and discuss with different practitioners to build a rich mental model.

  5. Connect the dots. The most valuable insights come from connections between topic areas. How does ontology design affect security? How does testing theory inform prompt engineering? These cross-cutting insights are where the SSA role becomes most powerful.

If you are new to most of these topics, consider this sequence:

  1. Start with knowledge representation and ontologies -- this is the foundation everything else builds on.
  2. Move to prompt engineering and context design -- this connects the meaning layer to the execution layer.
  3. Study multi-agent systems -- this is where individual agents become collaborative systems.
  4. Add software quality and testing -- this ensures your systems actually work.
  5. Layer in information security -- this protects what you have built.
  6. Incorporate product management and site reliability engineering -- these bring your systems to production.
  7. Weave ethics throughout -- ethical reasoning applies at every stage.

Staying current​

The field of AI systems design evolves rapidly. Here are strategies for staying current:

  • Follow the primary sources. Read documentation and technical reports from the teams building the models and frameworks you use. These are updated more frequently and more accurately than secondary sources.
  • Build and test regularly. The best way to understand new capabilities is to build with them. Set aside time each month to experiment with new tools, models, or techniques.
  • Join practitioner communities. Peer learning accelerates growth. Find communities where practitioners share real-world experiences, not just theoretical discussions.
  • Track the failure modes. Pay attention to incident reports, post-mortems, and case studies of AI systems that failed. Failure analysis teaches more than success stories.
  • Revisit fundamentals periodically. As the field evolves, revisit your understanding of fundamentals. New developments often reveal deeper implications of concepts you thought you understood.
  • Maintain a learning journal. Each week, write down one thing you learned, one thing you are confused about, and one thing you want to explore next. This simple practice creates compound learning over time.

Weekly self-check​

Use these questions at the end of each study week:

  1. What new concept did I learn about modeling or architecture?
  2. What can I now measure or evaluate that I could not before?
  3. What new risk or failure mode have I identified?
  4. What architectural or design decision have I made or revisited?
  5. What connection between topic areas did I discover?
  6. What question do I still not have a good answer to?