iSAQB® AGENTA - Architecture for Agentic Software Engineering Contexts

The curriculum uses this chapter to bring along participants with no prior AI experience.

• Separating AI-assisted, augmented, agentic and vibe coding
• LLMs as probabilistic tools: tokens, context, reasoning, knowledge cutoff
• The agentic loop as the ReAct pattern, and what extends it
• Autonomy as a spectrum, from spec-driven development to agent swarms

This is where a decision becomes an experiment instead of a meeting.

• The lifecycle of an architectural decision, and what agents accelerate in it
• Decomposing quality goals top-down into architectural drivers
• Finding structural pain points bottom-up, through analysis and monitoring
• Framing decisions as testable hypotheses and validating them with spikes

The largest block: what an agent knows decides what it builds.

• Which decisions, constraints and system boundaries an agent actually needs
• Comparing representations: ADRs, architecture rules, context files, diagrams as code
• Delivery at the right moment, through retrieval augmented generation, skills and the Model Context Protocol
• Why more knowledge is not better: contradictory context degrades the output

Knowledge alone is not enough; this chapter builds the enforcement.

• Preventive and detective controls, and their trade-off between precision and coverage
• Atomic structural rules against holistic quality attributes, thought of as fitness functions
• Where in the process each control belongs, when drift outpaces the review cycle
• Introducing controls into existing code incrementally, and improving them from feedback

The direction reverses here: from the system back to the documentation.

• Telling as-is from to-be, and human-facing from agent-facing
• Capturing decisions made across long agentic loops in ADRs
• Extracting technologies, patterns and business rules from code, tests and configuration
• Generating architecture views and keeping them versionable as Mermaid or Structurizr

This chapter settles what your company has to be able to evidence.

• The EU AI Act, copyleft licences and the roles in data protection
• Traceability through structured logging, prompt versioning and citation
• Billing models, vendor lock-in and data sovereignty when choosing a provider
• Prompt injection, tool misuse and sandboxing against the OWASP LLM Top 10

To close, the subject is your own role and your team's.

• Responsibility stays with people, and how you practise analytical skills deliberately
• Spotting bias in training data and in your organisation's own artefacts
• Measuring AI adoption hypothesis-driven, for example against DORA metrics