Code Smarter | .Net Meetup Karlsruhe January 2026

Transcript

1. From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices Daniel Sogl daniel.sogl@thinktecture.com Consultant | Speaker
2. Daniel Sogl • Consultant @ Thinktecture AG • MVP – Developer & Web Technologies • Focus: Developer Productivity & Generative AI • Socials: https://linktr.ee/daniel_sogl About me From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
3. Democratization of Software Development • Software development is now accessible to everyone – powered by natural language • GitHub Copilot: 50,000+ organizations • Path to "a billion programmers” - GitHub CEO The AI Coding Revolution https://www.elitebrains.com/blog/aI-generated-code-statistics-2025 https://www.netcorpsoftwaredevelopment.com/blog/ai-generated-code-statistics https://blog.github.com From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
4. 84% of Developers Are Already Using AI Tools • 84% using or planning to use AI coding tools (Stack Overflow) • 51% use them daily (Stack Overflow) • 85% regularly use AI for coding (JetBrains) • From pilots to production: 90% enterprise adoption predicted by 2028 (JetBrains) • Accenture: 90% report greater job satisfaction Adoption Has Exploded https://blog.github.com https://survey.stackoverflow.co/2025/ https://blog.jetbrains.com/research/2025/10/state-of-developer-ecosystem-2025/ From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
5. What the Data Actually Shows • 55.8% faster task completion (controlled study by GitHub) • Save 56 minutes per working day = 28 days annually (UK Gov study) • 8.69% increase in pull requests and 15% increase in PR merge rate (Accenture) • Acceptance rate: ~30% of suggestions (Google) Real Productivity Gains https://blog.github.com https://abc.xyz/site-map/default.aspx https://www.qodo.ai/reports/state-of-ai-code-quality https://www.gov.uk/government/publications/ai-coding-assistant-trial/ai-coding-assistant-trial-uk-public-sector-findings-report From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
6. The Perception vs. Reality – METR Study • Tool used: Cursor with Claude Sonnet 3.5 • Developer perception: Believed they were 20% faster • Actual performance: 19% SLOWER with AI tools • Perception gap: 40 percentage points difference The Reality Check https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
7. “We provide evidence that recent AI systems slow down experienced open-source developers with moderate AI experience completing real issues on large, popular repositories they are highly familiar with.” https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/ From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
8. Dangerous Over-Confidence • 75-80% believe AI code is MORE secure than human code • Yet 56% admit AI introduces security issues "sometimes or frequently" • Only 10% scan AI code for vulnerabilities • 79% bypass security policies to use AI tools The Trust Paradox https://apiiro.com/blog/ai-generated-code-security/ https://arxiv.org/html/2402.07632v1 From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
9. 10x Increase in Vulnerabilities in 6 Months • 45% of AI-generated code fails security tests • Java security failure rate: 72% • Secret leakage: 40% higher in Copilot repos • Privilege escalation: +322% • Design flaws: +153% • 86% fail to defend against XSS attacks The Dark Side - Security Crisis https://www.veracode.com/resources/analyst-reports/2025-genai-code-security-report/ From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
10. “Never Seen This Much Debt Created So Fast” • Code duplication: 8x increase in 2024 • Code churn climbing to 7% by 2025 • Code refactoring approaching zero • Copy-paste exceeds reuse Technical Debt Explosion https://www.gitclear.com/ai_assistant_code_quality_2025_research From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
11. 1. Confirmation Bias • AI writes what we expect to see • Confirms our existing assumptions • We don't look for errors we don't anticipate 2. Cognitive Load Shift • Before: Focus on implementation • Now: Focus on validation (much harder!) • Review requires more expertise than writing 3. Overconfidence in "Smart" Tools • 75-80% believe AI code is more secure Reality: AI has zero security awareness • Anthropomorphization: "Claude understands this” 4. The Measurement Problem • We measure wrong metrics (Lines of Code) • Not measured: Time for bugfixes, refactoring, code review • Survivor bias: Failed AI projects don't get published Four Reasons for the Trust Paradox From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
12. The Methodology Gap That's Killing Productivity What doesn't work: • "Vibe coding" without structure • No architecture review • Treating AI like autocomplete • Generating code without specs What works: • Custom instructions • Reusable prompts • MCP servers • Spec-Driven Development The Real Problem Isn't AI - It's How We Use It From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
13. Understanding LLMs for Developers From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
14. • Not all LLMs are equal: Performance varies widely • Understand reasoning vs. non-reasoning models • Context size matters greatly in coding scenarios • Benchmarks help guide practical model choice Choosing the Right LLM for Your Tasks From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
15. Model Provider Context Window SWE-Bench $ 1M I/O Token Claude 4.1 Opus Anthropic 200k tokens 74 % $15 / $75 Claude Sonnet 4.5 Anthropic 200k tokens 77 % $3 / $15 GPT-5.1 (high) OpenAI 1m tokens 76 % $1.25 / $10 GPT-5.1 Codex OpenAI 400k tokens 76 % $1.25 / $10 Gemini 2.5 Pro Google 1m tokens 63 % $1.25 / $10 Grok 4 xAI 128k tokens 69 % $3 / $15 Grok Code Fast 1 xAI 256k tokens 70 % $0.2 / $1.50 Popular LLMs for Software Development From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
16. • AMD tested 20+ local models — only a few are reliable for coding tasks • RAM is the real bottleneck — determines which model tier you can run • GGUF (Windows/Linux/Mac) vs MLX (Mac) – choose your model format wisely • Smaller models (<30B) fail at tool use, code reasoning & autonomy Local LLMs Source: https://cline.bot/blog/local-models-amd From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
17. • We can define so called custom instructions to provide extra context to all Copilot actions • A custom instruction is a markdown file with rules, instructions and guidelines defined by developers • Custom instructions can be added fine granular to specific file types and tasks • Custom instructions are part of a project and can be shared with all other developers • This pattern can also be used with Cursor, Windsurf, Claude Code or Gemini CLI • Supported by Copilot Plugin in VS-Code, VS, JetBrains How to solve that problem? From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
18. Custom Instructions: Example From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
19. Copilot knows two types of instruction files • Global instructions: copilot-instructions.md • Fine tuned instructions: custom.instructions.md • Instructions can be scoped to be included for defined file types like *.ts, *.html, *.css • Instructions can also be scoped for Copilot specific tasks like generating commit messages, generating tests or reviewing code Tip: Keep your instructions short. Each condition will be part of your context window Custom Instructions hierarchy From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
20. Prompt files are reusable prompts for tasks like code generation or code reviews • The content is defined in Markdown files • Prompt files are standalone and can be run directly in the chat • Example: A security review prompt file references general security practices and adds specific instructions on how to report findings Two scope types in VS Code: • Workspace prompt files: Available only within the workspace, stored in .github/prompts • User prompt files: Available across multiple workspaces, stored in the current VS Code profile Tip: Prompt files can use instruction files to reuse common guidelines and add task-specific instructions Prompt Files From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
21. Prompt Files: example From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
22. • Open-source protocol to integrate AI tools with external sources • Developed by Anthropic • Client-Server architecture: AI applications (clients) request context from external services (servers) • Security: secure data handling on your own infrastructure • MCP Server transport: local standard input/output (stdio), server-sent events (sse), and streamable HTTP (http) for MCP server transport What is MCP? From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
23. Tool Calling lets LLMs use external functions - LLMs generate structured requests to execute specific tools when they need capabilities beyond their training Extends capabilities - Enables real-time data access, performing actions, and running computations the LLM can't do alone Common examples - Web search, calculators, APIs, databases, code execution, email, and calendar systems Makes LLMs current and actionable - No longer limited to training data; can fetch live information and complete multi-step tasks Excursion – Tool/Function Calling From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
24. • Workspace settings: add a .vscode/mcp.json file in your workspace to configure MCP servers for a workspace and share configurations with team members • User settings: specify the server in your user settings to enable the MCP server across all workspaces • Automatic discovery: enable auto discovery of MCP servers defined in other tools, such as Claude Desktop • Tool executions have to bee confirmed and can be auto confirmed if wanted Using MCP in GitHub Copilot From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
25. • Claude Desktop supports MCP Servers • Remote Servers can be configured through the web dashboard • MCP Server can be configured the same way as with VS Code • Built in support for: GitHub, Google Services, Slack, Notion, Sentry and more • MCP makes Claude Desktop the most powerful AI companion app for non developers, compared with ChatGPT Desktop Claude Desktop: MCP Setup From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
26. Cyclomatic Complexity: measures how complex a code is based on the number of decision points Code Churn: measures the code added, modified, or deleted over time Code Coverage: quantifies the percentage of your codebase covered by automated tests Code Security: measures how resistant the code is to attacks and risks Code Documentation: measures the amount and quality of the documentation accompanying the code Code Duplication: measures the amount of code repeated or copied in different parts of the codebase Code Bug Issues: measures the number of bugs or defects found in the code per unit of code size Code Quality Metrics From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
27. • Adding static code analyses can track your code quality without adding more complexity to your project • Tools prevent adding complex, buggy or duplicated code into your existing code base on each commit or pull request • Many tools are free to use for public projects • Most tools support all common programming languages • Quality gates can be configured company wide Static Code Analyses From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
28. • Language specific: ESLint, SonarLint, pylint, Qodana • Language agnostic: SonarQube (Cloud), Codacy • Security: Snyk, Veracode Tools From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
29. • AI assistants (Copilot, Claude, Gemini) will be as standard as Git & CI/CD • Boiler-plate work shrinks → focus shifts to architecture, design & product insight • New responsibilities: AI coach, safety reviewer, data-governance advocate • Human strengths remain irreplaceable: creativity - context - empathy The Developer Role Is Evolving From Copilot to Autopilot Integrating AI Assistants into Modern Software Engineering Practices
30. Thanks for your attention Daniel Sogl @sogldaniel daniel.sogl@thinktecture.com