Skip to content

Armin Ronacher (mitsuhiko)

Austrian developer; creator of Flask and Jinja2; former long-time engineer at Sentry (~10 years, left April 2025 to build something new). Blog: lucumr.pocoo.org. Now a heavy contributor to Pi and one of the most-quoted 2026 voices on why the output-maximalism mode of AI coding is dangerous.

Disposition

  • Grounded skeptic of the prevailing "agent maximalism" narrative even while shipping agent-native tooling himself.
  • "The companies claiming that all of their code is now written by agents — we know the quality is garbage, we feel it in our bones when we use your product."
  • Values human judgment firmly at the heart of engineering work; not anti-agent, anti-outsourcing-thinking.

Key claims

  • Slow the F down. Blog-post thesis and the gravitational center of this episode. See slow-the-f-down. Core math: 10× code output × half your error rate = still 5× more errors shipped. Multiply that by the dark-factory fantasy of 100 agents and the codebase's rate of deterioration compounds past any single reviewer's throughput.
  • Non-engineers writing code is genuinely new and genuinely a problem. A PM vibing a feature into existence is useful as a spec-generating artifact, not as something to merge. Reviewer burden has tipped from "was this right" to "do I understand what they were even trying to do," which is psychologically worse.
  • Complexity down is the one codebase metric that still matters. Rejecting features that don't fit the current architecture is the lever. This is the opposite of the industry's "burn as many tokens as possible" posture.
  • MCP is the wrong abstraction. See case-against-mcp. The agent is extremely good at running code; MCP is an opaque RAG-like middle layer the model doesn't really see. His speculative "MCP2" is OAuth + generated SDK libraries (Stainless-style) + direct HTTP against OpenAPI specs — i.e. let the model write code against ordinary APIs.
  • Prompt-request ≠ pull-request, but the prompt alone isn't the deliverable. A terrible implementation is more informative than a polished prompt because it pins the intent. Disagrees with peter-steinberger's stronger "just send me the prompt" position. See prompt-request-over-pull-request.

Workflow signature

  • Vibed a kids' game with Pi end-to-end; wanted it to not look like slop, treats personal taste as a review filter.
  • Heavy user of terminal-native coding (CLI > MCP > GUI). The agent edits his own files; giving it direct filesystem access was the moment it clicked ("that was it for me").
  • Uses Pi extensions sparingly — a couple for GitHub-URL widgets, that's it. Treats most extensibility as temptation toward complexity.

Cross-references