Course Outline
From autocomplete to agents: why agents fail
• Anatomy of a coding agent: model, harness, tool surface, context, permissions.
• Where each tool sits: Claude Code, GitHub Copilot, Cursor, Codex CLI, Gemini CLI.
• A taxonomy of failure: wrong context, wrong tools, no feedback, unbounded autonomy.
Demonstration: The same task, run well and run badly, side by side.
Context engineering
• The context window as a budget: what earns a place in it.
• AGENTS.md, CLAUDE.md, .cursor/rules, copilot-instructions.md — one concept, several filenames, one source of truth.
• Conventions, build and test commands, architectural boundaries.
• Retrieval versus explicit context; task decomposition and sub-agents.
Lab: Write repository context for an unfamiliar Python service, then re-run a failing task and compare the output.
Reusable workflows and Agent Skills
• Choosing the abstraction: instruction file, skill, custom command, or plain script.
• Anatomy of a skill: triggering, instructions, bundled scripts, progressive disclosure.
• Portability across tools, and where lock-in begins.
• Versioning, review, and distribution across a team; common anti-patterns.
Lab: Build and test a reusable workflow that enforces a house coding standard.
MCP: connecting agents to real systems
• Architecture: clients, servers, tools, resources, and prompts; stdio and HTTP transports.
• Servers that earn their place: Git hosting, issue trackers, databases, browsers, internal APIs.
• When a CLI or a script beats an MCP server.
• Tool-surface hygiene: why more tools means less reliability.
Lab: Wire up MCP servers and take a ticket end to end — issue, branch, patch, tests, pull request.
Feedback loops and evaluation
• Tests, types, and linters as the agent’s ground truth; test-first work as a control mechanism.
• CI as the outer loop, and review discipline for agent-authored diffs.
• Golden-task evaluation sets: what to measure and how to catch regressions.
• Cost and latency as first-class metrics.
Lab: Build a small evaluation set and score two agent configurations against it.
Security and guardrails
• Prompt injection through issues, pull requests, READMEs, dependencies, and fetched pages.
• Permission models: allowlists, approvals, read-only tools, network egress control.
• Secret hygiene and sandboxing: containers, ephemeral credentials, limiting blast radius.
• Supply-chain risk in third-party MCP servers and shared skills.
Lab: Watch an agent get hijacked by a poisoned repository, then harden the setup so it does not.
Rolling this out to a team
• A staged adoption path; what to standardize and what to leave to individuals.
• Metrics that indicate real value, and the ones that do not.
Requirements
• Working knowledge of Python, Git, and the command line
• Some prior exposure to an AI coding assistant
• NobleProg will set up Dadesktop VMs for participants with Docker, VS Code, and Python 3.11 or later
• A working AI coding assistant of the participant’s choice: Claude Code, GitHub Copilot, Cursor, Codex CLI, or Gemini CLI. Labs are tool-agnostic, and instructions are provided for each platform.
Audience
• Software engineers, tech leads, and architects who use AI coding assistants but struggle to achieve reliable results
• Platform and developer-experience engineers rolling out AI tooling across teams
• Engineering managers establishing standards, guardrails, and success metrics.
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
