Agentic Coding

Transcript

1. Agentic Coding: Beyond Prototyping Sebastian Korfmann Tech Strategy Consultant - Founder of Hekto.dev
2. "Artisanal coding is like hand- tilling a commercial farm— beautiful, but not how we feed the world"
4. Terminal vs IDE "there's a good chance that by the end of the year, people aren't using IDEs anymore... we're trying to avoid over investing in UI... given how models are progressing, that might not be useful work" (in a Q&A session at the "Code with Claude" conference, May 2025) Anthropic’s Boris Cherny (Claude Code)
5. What is Claude Code? An AI coding agent that works in your terminal (or anywhere, really) • Builds up context on demand • Leverages Prompt Caching • Have seen it work for more than 20 minutes • Limiting factor: Primarily available context and tools npm install -g @anthropic-ai/claude-code
6. Example: Bluesky Frontend App • First attempt of using Claude Code • Entirely driven through Claude Code • A couple of hours experimenting
8. Example: Soccer Analysis Experimented during DFB Pokalfinale • Wondered if I could build real-time commentary • Set it up in like 2 hours • Watched the game with AI analysis Cost me about $1.25 in Gemini API calls for the Video Analysis
10. Time to Reflect starting to fully buy into the idea of agentic coding
11. Claude Code • For both experiments (plus a few more shorter ones): ~ 85 EUR in API Cost • Decided to go for the MAX Plan of claude.ai • That's 200 USD / month
12. Less constraints Stopped being concerened about cost
13. Example: A Miro Board via MCP • Used Claude Code with MCP (Model Context Protocol) integration • Direct API control of Miro • About 10 minutes from text to board Shows how tools can connect in unexpected ways
15. Example: Mobile Development Tried coding from my phone • Used WebSSH and Tailscale • Built a React app on the sofa • Deployed to Vercel Return of the thin client?
18. Example: Claude Code Open Telemetry • Rails App • Postgres • TimescaleDB
19. Example: Hekto.dev • Elixir Phoenix • MCP Gateway • RFC driven • Midsized Codebase • Runs in Production
20. Other Tools in This Space Different approaches, similar goals Claude Code - Terminal-based approach Goose (Block) - Desktop app with project generation Sketch - Browser-based with containers Cursor - IDE integration Each has its own philosophy and tradeoffs
21. An Interesting Observation We're only scratching the surface Example: Anthropic's designers now ship production code "Large state management changes you typically wouldn't see a designer making"
22. What I've Learned Some practical insights
23. Planning Makes a Difference Write specs together When you write RFCs or design docs with AI: - Makes a lot of difference vs single shot prompting - Helps LLMs as it helps humans - Catches edge cases early - Suggests alternatives you hadn't considered Example: Spend 10 minutes on an RFC, save way more time down the road
25. Embrace the Feedback Loop • Find some form of automated feedback loop early on • Tests, Compiler, Browser (MCP), ...
26. You're still responsible vibe coded, rfc driven, thoroughly tested? Whatever, but you're still responsible for what you're shipping :)
27. About Costs Some real numbers I tracked my usage with OpenTelemetry: - One heavy coding day: ~$175 worth of API calls - Claude Max subscription: $200/month The subscription model makes more sense for regular use.
30. The Economics Quick comparison Fully loaded Senior Engineer (Germany): ~€10,000/month Working days: ~20 Daily cost: €500 Claude Max: €200/month If it saves you half a day per month, it pays for itself. People report much more than that.
31. How Teams Might Change Some observations • smaller teams, larger scope • everyone ships to production • cost of prototyping approaching zero • way more software can and likely will be created • SaaS models? Still early days, but the impact will be massive
32. Looking Forward The trajectory is clear 2019: GPT-2 struggled with basic tasks 2022: ChatGPT / Copilot wrote code snippets 2024: Cursor builds full fledged features 2025: Agentic Coding 2026: ? The improvements have been consistent and exponential.
33. AI Task Completion is Doubling Every 7 Months The length of tasks that AI can successfully complete has been growing exponentially, doubling approximately every 7 months over the past 6 years.
34. Common Concerns Let's address the elephants "AI can't handle complex codebases" Actually, it does quite well with large projects "What about code quality?" In my experience, it writes thorough tests and catches edge cases "Should we wait for better models?" Current tools are already quite capable
35. Would you drive a car with agentic code? • what makes you so sure that human code is better, they make errors all the time • it's all about having testing systems in place, human code or not
36. Read also https://fly.io/blog/youre-all- nuts/
37. Takeaways • Don't wait, it's good enough already • Code is not an asset, it's a liability • Apply to uncommon use-cases • Actual limit: your imagination and access to tools • You get what you pay for
38. Getting Started If you're curious # Install Claude Code npm install -g @anthropic-ai/claude-code # Or try another tool like Goose, Sketch, or Cursor # Pick a small project to experiment with Start small, see what works for you.
39. Final Thoughts This technology is moving fast. What's not working today, might in 3 months. The key is to start experimenting and learning. Thanks for listening!
40. Questions & Discussion Sebastian Korfmann • @skorfmann.com (Bluesky) • sebastian@korfmann.net • hekto.dev Happy to chat about experiences, concerns, or ideas.
41. Links • https://bsky.app/profile/skorfmann.com/post/ 3lpt4zqqygs2d • https://bsky.app/profile/skorfmann.com/post/3lqpi2an7qs2i • https://bluesky-client-2oskmsae8-hekto.vercel.app • https://www-cdn.anthropic.com/ 58284b19e702b49db9302d5b6f135ad8871e7658.pdf • https://metr.org/blog/2025-03-19-measuring-ai-ability-to- complete-long-tasks/