Agentic Coding

Introduction to Agentic Coding

MLS · Machine Learning Systems · Workshop
From vibe coding to engineering with AI agents

🤖 Agent Loop · 🔌 MCP · 🔧 Tools · 🧠 Memory · 📐 Skills

---

The Road to Agentic Coding

When	What	Note
2023	“The hottest new programming language is English.”	— Karpathy
Feb 2025	“Vibe coding” coined	“Forget the code exists.” Collins Word of the Year.
Late 2025	The hangover	Security flaws. Unmaintainable code.
Feb 2026	“Agentic engineering”	Orchestrate agents with oversight. Human as reviewer.

Key shift: From “accept all, don’t read diffs” → “define goals, review results, verify performance.”

---

The 8 Stages of Dev Evolution to AI

1. Zero/near-zero AI. Maybe code completions, sometimes ask Chat.
2. Coding agent in IDE, permissions on. Sidebar asks to run tools.
3. Agent in IDE, YOLO mode. Trust goes up, permissions off.
4. Wide agent fills the screen. Code is just for reviewing diffs.
5. CLI, single agent, YOLO. Diffs scroll by. May or may not read.
6. CLI, multi-agent, YOLO. 3–5 parallel instances. Very fast.
7. 10+ agents, hand-managed. Pushing limits of orchestration.
8. Building your own orchestrator. Automating the workflow itself.

Ref: Steve Yegge — “Welcome to Gas Town” (2025)

---

What Changed: Capability Leap

Early 2025 — Autocomplete

💬 Suggest lines → 👤 Review
         ↑                |
         └── 👤 Fix ← 👤 Test

Human does all the work.

2026 — Autonomous Agent

📖 Read repo → 🧠 Plan
       ↑              |
       |        ⚡ Code → 🧪 Test
       |                     |
       └──── 🔧 Fix ← 🔄 Loop

Agent loops until pass; human reviews.

---

Model Timeline

Date	Model	Highlights
Sep 2025	Sonnet 4.5	SWE-bench 77.2%, 30+ hour tasks
Nov 2025	Opus 4.5	Flagship reasoning, major coding gains
Feb 2026	Opus 4.6 + Sonnet 4.6	Agent Teams, 1M context, SWE-bench 79.6%

Key insight: Models went from “fancy autocomplete” to autonomous multi-step execution. That’s the difference between Copilot 2024 and Claude Code 2026.

---

The Agent Loop: Reason → Act → Observe

🧠 Reason ──→ ⚡ Act ──→ 👁️ Observe ──→ 👤 Human Review
    ↑                                        │
    └────────────── loop ────────────────────┘

• Reason — Plan approach
• Act — Write code, run commands
• Observe — Read test output
• Human Review — Approve & ship

Human role: Define goals, review results, understand the hardware. The agent writes code; you verify it’s correct and efficient.

---

Don’t just ask “do everything” — there’s a smarter way

Let’s learn how to break tasks into agentic skills.

---

From Agents to Skills: The Evolution

• 2024–25 — Simple Prompt Agents — save users time, reduce repetitive context input. Each conversation starts from scratch.
• 2025 — MCP (Model Context Protocol) — replaces ad-hoc API integrations. Gives AI structured documentation about tools + JSON-formatted output for precise tool calls.
• 2025–26 — Skills + CLAUDE.md — persistent project rules, coding patterns, test workflows. Context survives across sessions. Agentic programming can finally manage context at scale.

Result: Agent + Skills + Agent VM = an autonomous coding partner, not just a chatbot.

---

The Stack: Agent + MCP + Skills

🤖 Agent — the reasoning engine
Reason → Act → Observe → Loop. Drives the autonomous cycle.

🔌 MCP — connect external tools
Gives AI structured docs about external tools + JSON output for precise API calls.

📐 Skills — what to do & how
Folders of instructions (SKILL.md) + scripts + reference files. Agent discovers & loads them on demand.

Most skills execute locally but in separate virtual environments (e.g. Python venv, Node env) to isolate dependencies and side effects.

Skills encode workflows; MCP connects tools; the Agent reasons over both. Together they turn a chatbot into an autonomous coding partner.

Ref: Equipping agents for the real world with agent skills

---

Skill Example: Simple Skill

A single SKILL.md file with all instructions.

• Name + description — agent decides when to load
• Instructions — step-by-step workflow
• No extra files — everything fits in one document

Good for: commit conventions, code review checklists, deploy scripts.

Ref: Equipping agents for the real world with agent skills

---

Skill Example: Complex Skill

SKILL.md + optional bundled components, loaded on demand.

• SKILL.md — core instructions (always the entry point)

Common optional extensions:

• code/ — executable scripts, helpers
• templates/ — boilerplate, scaffolding
• forms/ — structured input schemas
• reference/ — docs, examples, specs

Agent only pulls each component when the task actually needs it — keeps the context window lean.

Ref: Equipping agents for the real world with agent skills

---

Comparing Agentic Tools

	Claude Code / Cursor	Copilot (Plugin)
Form factor	Terminal / Full IDE	VS Code extension
Context	@ to add files & folders	@ to add files
Commands	/ to invoke tools & skills	/ to load saved prompts
Agent loop	YOLO mode, git worktree for parallel agents	Chat + inline

• Claude Code / Codex — best-in-class context management. Auto-compresses long conversations. Built-in Plan Mode for orchestration.
• Other CLIs (Aider, OpenCode, etc.) — similar results with plugins like Taskmaster, but require manual config.
• IDE Plugins (Copilot, etc.) — IDE-scoped context makes it harder to scale to multi-agent. Cursor’s full-screen agent is a step in this direction.

Note: boundaries aren’t absolute — Cursor has its own agent mode, and Cline as a plugin also supports full agent workflows. What matters is learning to use agent skills effectively.

---

Tests = Agent’s Feedback Signal

Agent writes code
        ↓
   Run unit tests
        ↓
   ✅ pass → Next task
   ❌ fail → Fix → Re-run ⟲
        ↓ (all tests pass)
   Run integration / E2E tests
        ↓
   Profile & benchmark

Without clear pass/fail, the agent can’t self-correct.

---

Best Practices & Pitfalls

✅ DO

1. One task at a time → test → next
2. Feed errors + reference code to agent
3. Ask agent to explain its reasoning and key decisions
4. Profile and benchmark to verify performance
5. Use profiling tools to locate bottlenecks
6. Use CLAUDE.md for project rules

❌ DON’T

1. Generate entire project at once
2. Ship code you haven’t read
3. Ignore shape mismatches or type errors
4. Ask “why broken?” without the error message
5. Accept “optimal” without checking actual performance
6. Let the agent modify core infrastructure files without review

---

Quick Start: Get Going Today

1. npm install -g @anthropic-ai/claude-code — Install (Node.js 18+)
2. cd your-project && claude — Launch in your project directory
3. "Read the README and explain the project structure" — Agent analyzes
4. "Implement feature X following the existing patterns. Run the tests." — Start small
5. "Fix any failing tests and explain what went wrong" — Let the agent loop

Alternatives: Cursor · GitHub Copilot (free for students) · Antigravity

---

Summary

Agentic coding is not about replacing you

• 🔄 Agent loop: Reason → Act → Observe → Loop
• 🔧 Tools + Memory + Skills = capable agent
• 🔌 MCP standardizes agent ↔ tool connections
• ✅ Tests enable the autonomous feedback loop
• 📊 Profiling — locate bottlenecks, don’t guess
• 🧠 Understanding the code is still your job

Questions? 🚀