Building Reliable AI Coding Agents with Skills, MCP and Guardrails Training Course

From Autocomplete to Agents: Understanding Failure Modes

•     Deconstructing a coding agent: model, harness, tool surface, context, and permissions

•     Positioning of various tools: Claude Code, GitHub Copilot, Cursor, Codex CLI, and Gemini CLI

•     Categorizing failures: incorrect context, inappropriate tools, lack of feedback, and unbounded autonomy

Demonstration: Comparing the same task executed optimally versus poorly, side by side

Context Engineering

•     Treating the context window as a finite budget: determining what deserves inclusion

•     AGENTS.md, CLAUDE.md, .cursor/rules, copilot-instructions.md — a single concept with various filenames, serving as the source of truth

•     Defining conventions, build and test commands, and architectural boundaries

•     Retrieval versus explicit context; task decomposition and the use of sub-agents

Lab: Creating repository context for an unfamiliar Python service, then re-running a previously failing task to compare results

Reusable Workflows and Agent Skills

•     Selecting the right abstraction level: instruction files, skills, custom commands, or plain scripts

•     Anatomy of a skill: triggering mechanisms, instructions, bundled scripts, and progressive disclosure

•     Portability across different tools and identifying where vendor lock-in begins

•     Versioning, review, and team distribution; addressing common anti-patterns

Lab: Developing and testing a reusable workflow that enforces internal coding standards

MCP: Integrating Agents with Real-World Systems

•     Architectural components: clients, servers, tools, resources, and prompts; utilizing stdio and HTTP transports

•     Identifying valuable servers: Git hosting, issue trackers, databases, browsers, and internal APIs

•     Scenarios where a CLI or script is more effective than an MCP server

•     Managing tool-surface hygiene: explaining why an increased number of tools reduces reliability

Lab: Configuring MCP servers to manage a ticket end-to-end — from issue to branch, patch, tests, and pull request

Feedback Loops and Evaluation

•     Using tests, types, and linters as the agent’s ground truth; adopting test-first practices as a control mechanism

•     CI as the outer loop, and maintaining review discipline for agent-generated diffs

•     Creating golden-task evaluation sets: defining metrics and detecting regressions

•     Treating cost and latency as primary performance metrics

Lab: Constructing a small evaluation set to score two different agent configurations

Security and Guardrails

•     Mitigating prompt injection via issues, pull requests, READMEs, dependencies, and fetched pages

•     Implementing permission models: allowlists, approval processes, read-only tools, and network egress control

•     Secret management and sandboxing: utilizing containers, ephemeral credentials, and limiting blast radius

•     Assessing supply-chain risks associated with third-party MCP servers and shared skills

Lab: Observing an agent compromised by a poisoned repository, then hardening the configuration to prevent recurrence

Team Implementation Strategy

•     Defining a phased adoption path; deciding what to standardize and what to leave to individual discretion

•     Identifying metrics that reflect genuine value versus those that do not