Course Outline
Evolution from autocomplete to autonomous agents: Understanding failure modes
• Structural components of a coding agent: underlying model, execution harness, tool interfaces, contextual data, and permission scopes
• Operational positioning of key platforms: Claude Code, GitHub Copilot, Cursor, Codex CLI, and Gemini CLI
• Categorization of failure points: contextual misalignment, improper tool selection, lack of feedback mechanisms, and unchecked autonomy
Demonstration: Comparative analysis of task execution outcomes under optimal versus suboptimal conditions
Context engineering strategies
• Treating the context window as a limited resource: prioritizing high-value information inclusion
• Standardizing configuration files (AGENTS.md, CLAUDE.md, .cursor/rules, copilot-instructions.md) to establish a single source of truth for instructions
• Defining coding conventions, build and test protocols, and architectural boundaries within the context
• Balancing retrieval-based approaches with explicit context provision; implementing task decomposition and sub-agent structures
Practical Exercise: Developing repository context documentation for an unfamiliar Python service, followed by re-executing a failed task to evaluate improvements
Implementing reusable workflows and agent skills
• Selecting appropriate abstraction levels: instruction files, specific skills, custom commands, or standalone scripts
• Anatomy of a skill component: trigger mechanisms, instruction sets, bundled scripts, and progressive information disclosure
• Ensuring portability across different tools and identifying points of potential vendor lock-in
• Establishing version control, review processes, and team-wide distribution practices; identifying common anti-patterns
Practical Exercise: Creating and testing a reusable workflow designed to enforce internal coding standards
MCP: Integrating agents with enterprise systems
• Architectural components: clients, servers, tools, resources, and prompts; utilizing stdio and HTTP transport protocols
• Identifying high-value server integrations: Git hosting services, issue tracking systems, databases, browser automation, and internal APIs
• Determining when command-line interfaces or direct scripting offer superior efficacy over MCP server implementations
• Maintaining tool-surface hygiene: analyzing the correlation between expanded tool access and decreased reliability
Practical Exercise: Configuring MCP servers to manage the full lifecycle of a task, from issue creation to branch management, patching, testing, and pull request submission
Establishing feedback loops and evaluation frameworks
• Leveraging tests, type checking, and linters as ground truth for agent performance; adopting test-first methodologies as a control mechanism
• Utilizing Continuous Integration (CI) as an external feedback loop and enforcing rigorous review discipline for agent-generated code changes
• Developing golden-task evaluation sets: defining key performance indicators and detecting performance regressions
• Treating cost and latency as primary operational metrics
Practical Exercise: Constructing a baseline evaluation set and benchmarking the performance of two distinct agent configurations
Security protocols and guardrails
• Mitigating prompt injection risks originating from issues, pull requests, README files, dependencies, and external content retrieval
• Implementing robust permission models: allowlists, approval workflows, read-only tool access, and network egress restrictions
• Enforcing secret management and sandboxing practices: containerization, ephemeral credentials, and minimizing the blast radius of potential breaches
• Assessing supply-chain vulnerabilities associated with third-party MCP servers and shared skill libraries
Practical Exercise: Simulating an agent compromise via a poisoned repository, followed by hardening security configurations to prevent such incidents
Organizational implementation strategy
• Developing a phased adoption roadmap: determining which practices to standardize and which to leave to individual discretion
• Identifying key metrics that accurately reflect value delivery versus those that provide misleading signals
Requirements
• Proficiency in Python, Git, and command-line operations
• Prior experience with an AI coding assistant
• NobleProg will provision participant environments (Dadesktop VMs) with Docker, VS Code, and Python 3.11 or later
• A functional AI coding assistant of the participant's choice (e.g., Claude Code, GitHub Copilot, Cursor, Codex CLI, or Gemini CLI). Practical exercises are tool-agnostic, with specific instructions provided for each supported platform
Target Audience
• Software engineers, tech leads, and architects seeking to achieve reliable outcomes with AI coding assistants
• Platform and developer-experience engineers responsible for deploying AI tooling across teams for government
• Engineering managers tasked with establishing standards, guardrails, and success metrics for AI adoption
Testimonials (3)
The trainer is patient and very helpful. He knows the topic well.
CLIFFORD TABARES - Universal Leaf Philippines, Inc.
Course - Agentic AI for Business Automation: Use Cases & Integration
Good mixvof knowledge and practice
Ion Mironescu - Facultatea S.A.I.A.P.M.
Course - Agentic AI for Enterprise Applications
The mix of theory and practice and of high level and low level perspectives
