Section 01
Study Guide
Long-form notes and references.
This study guide focuses on the engineering principles and mental models required for designing robust AI agent systems, specifically tailored for the Claude Certified Architect — Foundations curriculum. The central thesis of this domain — referred to as the “North Star” — is that when an agent fails or behaves unexpectedly, the solution must be implemented in the Architecture, Orchestration, or Code layer, rather than the Prompt or Model layer.
North Star
When an agent fails or behaves unexpectedly, the fix belongs in the Architecture, Orchestration, or Code layer — not the Prompt or Model layer.
Glossary of Key Terms
- Adversarial Agent
- A specialized agent (Critic or Verifier) tasked with finding faults or hallucinations in the output of a Producer agent.
- Bounded Loop
- An agentic loop that includes a deterministic termination condition (like a turn counter) enforced by the orchestration code.
- Context Bloat
- The accumulation of irrelevant tokens in the context window, leading to high costs and Lost-in-the-Middle performance degradation.
- Deterministic Router
- A code-based or rule-based classifier that directs tasks to specific models or workflows without relying on a large LLM's reasoning.
- Exponential Backoff
- A resilience pattern where the system waits for increasing amounts of time between retries after a failure (e.g., a 429 rate limit error).
- Hub-and-Spoke
- An orchestration pattern where a central Coordinator delegates tasks to specialized subagents and manages the flow of information.
- Idempotency
- A property of an operation where multiple identical requests have the same effect as a single request, preventing duplicate actions.
- Jitter
- Random variations added to retry delays to prevent multiple agents from retrying at the exact same time and overwhelming a system (thundering herd).
- Liveness vs. Safety
- A distributed-systems principle; in agents, it ensures the system eventually produces a result (liveness) while never entering an invalid state (safety).
- Negative Quota
- A prompt instruction for a Reviewer agent that mandates finding a specific number of flaws, forcing the agent to be more critical.
- Saga Pattern
- A failure-management pattern for long-running transactions where a sequence of local transactions is coordinated to ensure system-wide consistency.
- Self-Healing Loop
- A design where tool errors are returned as structured data to the model, allowing it to read the error and attempt a correction.
- Stop Reason
- Metadata returned by the Anthropic API (e.g., tool_use, end_turn, max_tokens) that indicates why the model stopped generating text.
- Structured Trace
- A detailed, machine-readable log of every decision, tool call, and state transition made by an agentic system for the purpose of auditing.
- Typed Envelope
- A consistent JSON structure used for tool results that explicitly separates status, error categories, and content.