Home > 🎤 Presentations
🤖 Agentic Software Engineering
The New Paradigm for AI-Augmented Development
🎯 What is Agentic Software Engineering?
- AI agents that autonomously plan, write, test, and debug code
- Long-running sessions with tool execution
- Human oversight, not replacement
- New patterns, tools, and best practices
🌟 The Spectrum
| Level | Description |
|---|
| AI Prototyping | Fast, exploratory, prompt-driven |
| Directed Assistance | Constraints + patterns + success criteria |
| Agent Orchestration | Multiple parallel agents |
| Agentic Engineering | Persistent workflows + guardrails + accountability |
🔧 Core Patterns
- TDD for Agents - Write failing tests first
- Context Engineering - CLAUDE.md, instruction directories
- Just-in-Time Loading - Load context when needed
- Human-in-the-Loop - Confirm before destructive actions
| Category | Tools |
|---|
| Coding Agents | Claude Code, OpenAI Codex, GitHub Copilot, Cursor |
| Frameworks | LangChain, AutoGen, OpenClaw, CrewAI |
| Local | Ollama, LM Studio, Jan |
| Protocol | MCP (Model Context Protocol) |
📊 Benchmark Leaders
- SWE-bench: Claude Opus 4.6 (79.2%), Gemini 3 Flash (76.2%), GPT-5.2 (75.4%)
🔒 Security (OWASP Top 10)
- Sensitive Data Disclosure
- Tool Poisoning
- Memory Pollution
- Prompt Injection
- Unbounded Execution
📈 Key Trends
- Single agents → coordinated teams
- Long-running autonomous sessions
- Human oversight scales through collaboration
- Security-first architecture essential
- Cost management critical
🎯 Practical Next Steps
- Start with TDD
- Build your instruction directory
- Orchestrate, don’t micromanage
- Invest in observability
- Keep learning!
📚 Resources