Since many people have asked me lately what my Agentic Coding setup looks like, here it is.
Originally posted on LinkedIn on Aug 27, 2025
Since many people have asked me lately what my Agentic Coding setup looks like, here it is.
š¼ My Work/Time division:
70%: R&D of AI-powered tools for commercial business
20%: Engineering of state-of-the-art transformer-based neural networks on all kinds of data and scientific projects
10%: Research and reproduction of ML papers (wish it was more with the paper onslaught nowadays)
Hence I mainly use AI tools to build... more AI tools š
š¤ LLM Service Providers:
Claude Max plan 20x Pro (ā¬180/mo) - Includes regular Claude with Projects + virtually unlimited Claude Code
OpenAI Plus (ā¬20/mo) - Includes codex-cli, their coding agent similar to Claude Code
šØāš» Editor:
VS Code - Heavy with extensions, but Claude Code integration with Diffs works well
Command-line - just using claude [task] works great, cli works everywhere
š Setup:
Claude Web App: Set up Projects with documentation, voice transcripts, and rules. I develop ideas here, using web search to validate claims. For larger/research projects, I cross-pollinate ideas between Claude, OpenAI, and Gemini, adding my feedback on-the-fly.
Claude Code: The actual coding workhorse - still unbeatable among coding agentsCodex-cli: Code reviews using GPT-5-Thinking (Plus plan suffices since reviewing is less consumptive)
š ļø Project Structure:
ā CLAUDE.md - Project introduction and rules (Codex reads this too)
ā planning/
āā tickets/ - JIRA-style tickets with requirements, definition of done
āā reports/ - Implementation details after modifications by Claude Code
āā review/ - Codex reviews with filenames, code samples, specific issues
š Actual Flow:
1) Ticket Generation - Claude creates scoped, minimal JIRA-style tickets based on Project definition and my input
2) Development - Claude Code: branch ā code ā commit ā PR (supervised by me)
3) Code Review - Codex reviews, supervised by me
4) Amendments - Claude fixes issues based on review, generates change report
5) Re-review by Codex - If critical issues remain, repeat 4-5
6) Final Review - Critical review by Claude Code + my input
7) Merge - Back to step 1
I do not automate the flow completely - I read the generated code and can steer it in another direction if i do not like it. I seldom let Claude auto-accept changes.
So basically I run two terminals side-by-side in VS Code, 1 for Claude, 1 for Codex, and let them code and review turn-by-turn.
With new goodies and tools arriving every month, this workflow will definitely evolve. But I'm quite happy with the current setup - having both Claude and GPT-5 bicker and refine code is quite assuring to see.