# Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC Berkeley

## Executive summary

The presentation uses Anthropic's Claude Certified Architect (CCA) exam structure to guide engineers through agentic AI best practices by focusing on anti-patterns. Key principles include mastering 'agentic loops' over simple sequences, ensuring specialized subagents only have limited tools, and rigorously managing context via techniques like forking and compaction to prevent groupthink and token overflow.

## Key takeaways

- Prioritize Looping Mechanisms: Agentic workflows should leverage loops (e.g., `while true`) rather than simple sequences, as looping provides the necessary power for complex computation, paralleling the foundational concepts of Turing completeness.
- Use Stop Reasons for Control: Instead of blindly using an agent's response, robust workflows must branch logic based on the model's `stop reason`. This is crucial because the LLM cannot execute tools; it only provides parameters that external code must run.
- Specialize Agents and Tools: Avoid 'overloading' a single agent with all available tools. Specialized subagents, each equipped with one or two specific tools, are significantly more effective and maintain focus.
- Manage Context Isolation: To prevent context pollution and 'groupthink,' isolate subtask outputs (e.g., using a 'context fork') so that only necessary summaries return to the main thread, preventing unbounded context growth.

## Technical details

- Agentic Architecture & Anti-Patterns: Anti-patterns include: (1) Calling the model and using its response directly without checking the `stop reason`. (2) Loading one agent with every available tool. (3) Allowing subtask outputs to dump into the primary context, leading to context bloat.
- Claude Code & Context Management: Anthropic recommends a hierarchical structure for rules using the 'Claude MD file' (at project and directory levels) to control system responses. For large contexts, implement compaction algorithms when token counts exceed thresholds (e.g., 150,000 tokens).
- Multi-Agent Systems: When designing multi-agent systems, only pass the necessary 'claim and evidence' to a critic agent, deliberately withholding the reasoning process that generated them to mitigate groupthink.
- Batch Mode Execution: For cost efficiency, utilize batch mode for processing work. This can reduce token costs by up to 50% if the results are acceptable within a 24-hour delay.

## Practical implications

- When designing agent workflows, treat the LLM as a parameter extractor, not an executor. All tool calls must be wrapped in external code that checks the model's `stop reason` for successful execution flow.
- Implement context management layers (e.g., 'context forks') to ensure subtask outputs are summarized and compacted before being merged back into the main thread, preventing token overflow and confusion.
- Adopt a specialized micro-agent architecture where each agent is limited to a small set of tools relevant only to its function.

## Topics

Agentic AI, LLM Orchestration, Context Management, Design Patterns (Anti-Patterns), Claude Code Integration, Workflow Automation, Claude Certified Architect exam, Anthropic ecosystem

Source: https://www.youtube.com/watch?v=Z-c11pV_uvU