State of Agentic Coding

Ben and Armin close season 1 with AI employees that work nine to five and a debate about how to work effectively with coding agents. From software factories and product complexity to cloud development and agents that “optimize” away your debugging tools, they explore why cheaper code doesn’t make human judgment optional.

Geoff Huntley, creator of the Ralph Wiggum loop, joins Armin and Ben to discuss what comes after handwritten code. From DHH’s Rails World keynote and Pi Durable to a live self-modifying Lisp experiment, they explore how AI is changing software development – and question whether token economics, developer habits, and actual user demand can keep up.

Armin and Ben discuss AI watermarking, model-provider politics, open-weight models convergence, chat coding agents, and what's going on with GitHub. They also explore the tension between vibe coding and engineering craft, and what truly agent-friendly programming might look like.

Agent autonomy, provider lock-in, and what coding agents actually cost once the subscription stops hiding it. Plus the terminal's comeback and the incentives behind free tokens.

Recorded in Vienna, Ben and Armin sit down with Mario Zechner to vibe-check Fable and GLM-5.2, unpack how RL and agent harnesses normalize model jank, and ask whether newer models

are quietly getting worse. They also debate agent loops, trust and ROI, the rising cost of tokens and compute, and how to stay sane amid AI FOMO.

Ben and Armin mark one year of coding with agents, asking what has actually changed in real developer workflows — from terminal-based agents and prompt loops to the end of cheap token subsidies.

What happens when AI coding stops being subsidized? Ben and Armin dig into rising compute costs, security slop, acquisition strategy, GitHub cracks, and why coding traces may be the most valuable asset in the stack.

Armin and Ben discuss the death of the IDE, the growing software quality problem, AI-discovered vulnerabilities, slowing down, and more.

Armin and Ben discuss agentic sandboxes, the "old new models" (Opus 4.6 and ChatGPT Codex 5.3), vibe side projects, and slop forks.

Armin and Ben revisit their January predictions, discuss Openclaw as a personal coding assistant, and cover a range of topics including Opus 4.6 fast mode, LLMs building complex projects like browsers and compilers, the unhealthy state of the industry, and more.