saved
State of Agentic Coding #10 with Armin and Ben
Hraness wrote this summary from a saved copy of the source. Quotations are taken word for word from the source.
gist
Armin Ronacher and Ben Vinegar's 10th State of Agentic Coding episode covers AI watermarking, Cursor losing OpenAI models, cloud agents, and GitHub. Ronacher, who founded Earendil and builds the Pi coding harness, says model-written text is detectable even without Anthropic's watermark, and that Pi has started removing tests because the suite takes too long to run. Vinegar, co-founder of Modem and author of the Hunk diff tool, runs six checkouts across harnesses and models behind coverage, terminal-integration, and benchmark gates. Both say agent output still needs close architectural scrutiny.
ideas
- Ronacher says model-written text is detectable without a watermark. He points to detectors such as Pangram, trained on model output, and to one model copy-editing many writers so that its word choices stand out.
- The hosts compare OpenAI pulling its models from Cursor with Anthropic cutting off Windsurf. They tie the Cursor change to its acquisition by SpaceX and recount the Windsurf deal that ran from April to July 2025 and fell apart.
- Open-weight models fill gaps where subscription models refuse. Ronacher starts reverse-engineering sessions with Kimi before switching to an OpenAI model, and Vinegar moved a Hunk security fix to another model after the model he was using refused it.
- Coding agents are spreading to new surfaces rather than new product categories. The hosts cite Linear, Devin in Slack, and hosted cloud agents, and Ronacher says the coding agent is the only AI product he kept using in 2026.
- Vinegar merges agent code behind automated gates, while Ronacher has not found a responsible way to let go. Hunk has 94 percent coverage, a pseudo-terminal suite, and release-blocking benchmarks; Ronacher describes an agent shim that made a plug-in design work by discarding the design.
quotes
“saying no is materially harder if the machine keeps making pull requests.”
“nothing that I started using that's an AI product I still use and I tried so many.”
“if you're not having these kinds of experiences that's a bad sign.”
“I haven't found a responsible way for me to work with the AI, but I want to get there.”