Claude Certified Architect – Foundations

• Tool vs. Prompt Enforcement: Always prefer enforcing rules, execution orders (tool_choice), and business logic thresholds inside the schema or backend tool itself, rather than relying on prompt instructions.
• Context Budgeting: Prevent "lost in the middle" effects by aggressively trimming raw content and intermediate reasoning chains before downstream handoffs.
• Structured Data Over Prose: Prose collapses metadata. Always use structured JSON outputs to preserve attribution (via explicit IDs, dates, URLs) through multi-agent pipelines.
• Batch API Optimization: To save costs, chunk and resubmit only failed documents (using custom_id), submit batches based on strict SLA math + buffer, and refine your prompts on small samples before wide-scale batch execution.
• Tool Interfaces: Reduce agent guesswork. Use enums for bounded categories, require backend systems to handle ID-to-string lookups, and always provide explicit pagination controls instead of dumping massive datasets.
• Session Recovery: Never --resume blindly after a crash. Extract a structured checkpoint of completed work, inject it into a new session, and tell the agent exactly where to start.
• Schema Descriptions: Tool and parameter descriptions in the schema are the primary mechanism an LLM uses to know when to use a tool and how to format its inputs.

"I chose the most intuitive answers based on my current experience building AI Agents and AI Harnessing systems also experimenting running claude code with Local llms etc.

It was a tricky paper. Where for most of the MCQs, several options could technically solve the problem in the question, but the real challenge was picking the solution that follows the best practices. And this is where software engineering experience comes handy.

The exam tests your AI proficiency, AI ethics, system design & critical thinking and user experience mindset."

Based on these reflections and the patterns in the exam, keep these tactical strategies in mind when sitting for the test:

1. Always look for the structural fix. The wrong answers heavily feature "improve the prompt" or "tell the agent not to do it." The correct answers almost always feature structural API changes, tool_choice, enum parameters, and structured JSON flags.
2. Beware of over-spawning. If a coordinator already holds the context, spawning a subagent is wrong. Subagents are for fresh exploration, not re-processing existing data.
3. Machine Identifiers rule. If an agent needs to perform an action on a specific item (a document, a game, a user), the correct architecture always uses an explicit ID (game_id, citation_id), never an ambiguous text string.
4. Assume self-reported metadata is untrusted. Especially for MCP servers, readOnly flags are just suggestions. Base security bypasses on system-level trust, not tool annotations.