Appearance
Phase 5 Interactive Flashcards: Harness Engineering & MCP
Master the architectural concepts of Phase 5 (Harness Engineering & Model Context Protocol) with interactive flashcards. Test your recall on the AI Harness concept, MCP Architecture, Build-Verify-Fix Loops, Loop Detection, Trace Failure Taxonomies, and Compulsory Operational Safety.
📚 Deck 1: The AI Harness Concept & Environment Onboarding
Master the hypervisor mental model and LocalContextMiddleware:
🗂️ Deck 1: AI Harness & Onboarding
Card 1 of 2AI Harness ConceptClick or press Space to flip ↺
What is an AI Harness and what architectural role does it fulfill?
AI Harness Concept • AnswerClick to flip back ↻
The AI Harness is the deterministic software system built around a foundation model. Acting like an operating system kernel or hypervisor, it handles environment virtualization, dynamic context injection, deterministic middleware hooks, rate limiting, and telemetry.
💡 Architect Takeaway: Treat the LLM as a CPU and the Harness as the OS kernel providing system calls and access controls.
📚 Deck 2: Model Context Protocol (MCP) Architecture & Consumption
Master JSON-RPC 2.0, FastMCP, MultiServerMCPClient, and zero-trust tool consumption:
🗂️ Deck 2: MCP Protocol & Consumption
Card 1 of 3Model Context ProtocolClick or press Space to flip ↺
How does the Model Context Protocol (MCP) eliminate the M × N integration bottleneck?
M clients × N services → M + N Open Protocol over JSON-RPC 2.0Model Context Protocol • AnswerClick to flip back ↻
Instead of building custom point-to-point glue code between every LLM client and every data source, MCP standardizes how servers expose Tools, Resources, and Prompts over stdio or SSE/HTTP using JSON-RPC 2.0.
💡 Architect Takeaway: MCP decouples AI agents from proprietary external API integration code.
📚 Deck 3: Aggressive Self-Verification & Build-Verify-Fix Loops
Master programmatic test gates, closed-loop feedback, and eliminating vibe-based coding:
🗂️ Deck 3: Build-Verify-Fix Loops
Card 1 of 2Build-Verify-FixClick or press Space to flip ↺
Why is 'vibe-based' code review by LLMs fundamentally flawed in production?
Build-Verify-Fix • AnswerClick to flip back ↻
LLMs suffer from the 'hallucination of competence': after writing code, autoregressive completion biases them toward claiming success. Without programmatic execution, syntax errors, missing imports, and logic bugs go undetected.
💡 Architect Takeaway: Strip the model of the authority to declare completion based solely on its own opinion.
📚 Deck 4: Deterministic Middleware Hooks (Loop Detection & Checklists)
Master action hashing, breaking Doom Loops, and PreCompletionChecklistMiddleware:
🗂️ Deck 4: Deterministic Middleware Hooks
Card 1 of 2Deterministic MiddlewareClick or press Space to flip ↺
What is a 'Doom Loop' and how does LoopDetectionMiddleware break it?
ActionHash = SHA-256(tool_name || canonical_json(args))Deterministic Middleware • AnswerClick to flip back ↻
A Doom Loop occurs when an agent repeatedly executes the exact same failing tool call. LoopDetectionMiddleware maintains a sliding window of action hashes. If an identical call repeats ≥ 2 times, it aborts execution and injects an assertive replanning directive.
💡 Architect Takeaway: Prevents runaway token burn by catching and breaking repetitive tool failures.
📚 Deck 5: Trace-Driven Failure Analysis & The Evaluation Data Flywheel
Master the 4 failure taxonomies and converting LangFuse failure traces into automated Pytest regressions:
🗂️ Deck 5: Failure Analysis & Flywheel
Card 1 of 2Failure AnalysisClick or press Space to flip ↺
What are the 4 primary agentic failure taxonomies identified via distributed tracing?
Reasoning | Tool Invocation | Context Insufficiency | Hallucination of CompletionFailure Analysis • AnswerClick to flip back ↻
1. Reasoning Failure: Flawed logic despite complete context and valid tools. 2. Tool Invocation Failure: Schema mismatches, malformed JSON, or missing parameters. 3. Context Insufficiency: Missing background information or codebase context. 4. Hallucination of Completion: Declaring success without satisfying verification gates.
💡 Architect Takeaway: Isolating failure taxonomy reveals whether to fix prompts, schemas, retrievers, or test gates.
📚 Deck 6: Compulsory Operational Safety (Rate Limiters & Circuit Breakers)
Master AIRateLimiter (RPM/TPM/USD), cost caps, and ToolCircuitBreaker states:
🗂️ Deck 6: Operational Safety & Circuit Breakers
Card 1 of 2Operational SafetyClick or press Space to flip ↺
Why must an AIRateLimiter enforce a multi-dimensional boundary (RPM, TPM, Budget USD)?
Operational Safety • AnswerClick to flip back ↻
Counting HTTP requests alone ignores token volume: one request might consume 15 tokens while another consumes 100k tokens. A production limiter checks RPM (concurrency), TPM (provider limits), and daily USD spend (preventing Denial-of-Wallet).
💡 Architect Takeaway: Multi-dimensional rate limiting prevents quota starvation and catastrophic runaway cloud bills.