Playbook
Tasteful Skills
“Tasteful Skills” argues that the best agent skills are not documentation or best-practice lists.
Back to Podcast Digest
State of Agentic Coding #6 with Armin and Ben
TL;DR
The agentic coding boom is running into real economics — Ben and Armin argue the “token maxing” era is fading as companies realize they can’t casually spend figures like $250,000 per engineer on API usage, code review, and agent workflows.
Security harnesses are making AI-found vulnerabilities impossible to shrug off — Armin points to tools like Sentry’s Warden and the “copy fail” Linux root-execution bug as proof that even if the writeups look like slop, the vulnerabilities are very real and increasingly easy to discover.
GitHub is showing strain just as AI agents are hammering it harder than ever — repeated outages, weak tooling for bot-heavy open source maintenance, and public frustration from people like Mitchell Hashimoto are making “leave GitHub” feel less fringe and more like the start of a shift.
Armin’s Pi story went from obsession to acquisition — after months of using, discussing, and building on coding agents with Mario Zechner, Arendelle formally acquired Pi, with a stated goal of being responsible stewards of an open-source coding agent that can also power products like Leos.
The xAI/Cursor deal makes more sense as a data deal than a code editor deal — Armin’s take is that xAI has GPUs but lacks high-quality coding traces, while Cursor has exactly the kind of human-plus-agent training data that frontier labs now view as strategically priceless.
The deeper fight is over who owns the traces and data exhaust from AI work — both hosts keep coming back to the same unease: coding traces, meeting transcripts, and product usage are becoming the new gold, and companies will be tempted to sell them long before users fully understand what they’ve consented to.
The Breakdown
AI Engineer Europe, Miami, and the value of the hallway track
The episode opens with a recap of AI Engineer Europe and Miami, which both hosts treat as the conference circuit that matters right now. Armin says Europe felt noticeably more grounded than San Francisco AI culture — older crowd, less hype, more balanced conversations — and notes that attendees specifically thanked him, Christina, and Mario for bringing some needed friction and realism to the conversation.
Hardware costs are creeping outward from GPUs to RAM to SSDs
They revisit an old prediction about compute getting more expensive and conclude: yes, it’s getting worse. Ben’s seeing high-density RAM and NVMe prices jump when shopping for local LLM hardware, and Armin connects it to AI demand, prompt caching, semiconductor dependencies like helium, and broader energy and supply-chain pressure — the whole stack is getting squeezed, not just GPUs.
AI security moved from theory to “oh, this gets root”
The conversation turns to whether frontier labs are holding back models because they’re too good at finding vulnerabilities, but Armin’s bigger point is that even current models are already enough. He cites the “copy fail” Linux privilege-escalation bug and Sentry’s Warden harness, which reportedly found around 100 vulnerabilities, as evidence that specialized agent loops are now making offensive security dramatically more automated.
Token subsidies are ending, and the bill is finally visible
This is the core mood shift of the episode: the incentives of labs and customers are diverging fast. Armin contrasts an OpenAI conference talk basically saying “don’t worry about token spend” with Mario Zechner’s efficiency-first mindset, while Ben points to products like Graphite moving from seat-based to usage-based pricing and watching monthly bills jump from roughly $150 to $800.
Pi’s path from side obsession to Arendelle acquisition
Armin tells the backstory of how Pi became part of Arendelle: Peter Steinberger got him and Mario hooked on coding agents last spring, they kept swapping workflows and ideas, and by late 2024 Mario was increasingly fixated on preserving the earlier, less-chaotic feel of Claude Code. Pi became both a practical coding harness and a building block for Arendelle’s other agent product, Leos, and Armin frames the acquisition as a long-game relationship move, not a sudden startup transaction.
Why xAI buying Cursor is really about training data
They briefly detour into the giant Cursor acquisition and laugh at how absurd the headline valuation sounds compared with something like Ford. But Armin offers the cleanest explanation of the episode: xAI has compute but not enough valuable coding data, Cursor has human-guided coding traces with a built-in reward signal like “did the user commit,” and those traces may be one of the most strategically valuable assets in AI right now.
GitHub feels shakier, and alternatives are no longer a joke
Armin is openly frustrated with GitHub outages, stale search indexes, weak PR/issue tooling, and the feeling that agent traffic is stressing a system that already lacks strong leadership. Mitchell Hashimoto’s decision to move Ghostty off GitHub feels significant to him in a way earlier moves didn’t, because once developers of that caliber leave publicly, alternatives like Pierre, Entire, Tangled, or even self-hosting stop feeling hypothetical.
Side projects: drawing for agents, and teaching kids with Pi
The lighter section is still revealing: Armin has been building a game with his kids and created a Pi extension that opens TLDraw so they can sketch ideas and send them back to the agent, while Ben built a terminal-first drawing tool to visually tell agents “no, put it here.” It’s a very on-brand ending: both of them are building tiny interfaces that add friction and clarity, because sometimes the fastest way to guide an agent is not more words — it’s a picture.
The closing note: reward the careful builders, not the reckless ones
The episode ends on a surprisingly earnest note about data abuse, consent theater, and irresponsible AI products shipping too fast. Ben says he wants the slow, painful, hard work of doing software responsibly to be rewarded again, and Armin folds that into Arendelle’s pitch: products like Pi and Leos should give users clearer model choice, clearer data handling, and a real alternative to the current “just click accept and hope” AI economy.
Share
Keep Reading
Two Ways to Follow
The Weekly Echo. The inbox-shaped summary of what mattered.
New editorials announced here.
