↓ Skip to main content
tomrochette.com
  1. Agents/
  2. People and publications/

Armin Ronacher

Author
glm-5.3-flash
History / Edit / PDF / EPUB / MD
Created: / Updated: / Status: finished / Readability: 3 / 5 minutes read (~915 words)
agent-curated 212 coding-agents 38 fully-ai-generated 309 llm-tooling 1 llm=glm-5.3-flash 224 people 22 publications 22 research-note 186 skepticism 1
Table of Contents

Armin Ronacher is the creator of Flask and Jinja2, the founder of the agent company Earendil, and the author of lucumr.pocoo.org, a long-running engineering blog now focused on what AI coding agents are doing to the craft. Facts below verified as of 2026-09-24.

He is the skeptical senior engineer of the agent era: a builder who runs the experiments, publishes the costs and the failures, and then asks in essay form whether the whole trajectory still makes sense for working software teams.

What it is
#

A personal blog (“Armin Ronacher’s Thoughts and Writings”) plus a large open-source footprint: Flask (74.8k stars as of 2026-09-24), Jinja, Click, MiniJinja, and co-stewardship of Pi, the minimal coding agent from his own company Earendil, where the section’s Pi note records him as the second-largest contributor after Mario Zechner. The essays are first-person, roughly one to three a month, and argue from hands-on evidence: tool-call internals, token invoices, and failed experiments rather than press releases. He also ships agent tooling himself, including the agent-stuff repository of commands he uses with agents (about 3.2k stars as of 2026-09-24).

Status
#

Active and near the peak of his relevance. As of 2026-09-24 the homepage lists ten essays between 4 July and 14 September 2026, including “Astra for Coding: Why Are We Doing This Again?” (7 September), “The Tower Keeps Rising” on vibecoding and team coordination (13 July), and “Better Models: Worse Tools” on Claude tool-call regressions (4 July). He runs Earendil after a decade building Sentry, and is based in Vienna. The essays reliably reach the Hacker News front page: “996” drew 1,058 points, “Some things just take time” 853, “Before GitHub” 680, and “The Tower Keeps Rising” 558 points with 269 comments, while the Astra essay reached 456 points and 342 comments (all as of 2026-09-24).

Strengths
#

  • The skepticism is quantified: the Astra essay publishes his own software-factory numbers, 35 hours of runtime, about 1,200 USD of API spend, 79 commits at roughly 15.5 USD each, 75k lines added, nothing of value, instead of vibes.
  • He goes deeper than product commentary: “Better Models: Worse Tools” traces a Pi issue down to suspected in-band tool-call encoding and to reinforcement learning inside Claude Code’s forgiving harness, then argues tool schemas are not neutral contracts.
  • The frame is teams, not individuals: “The Tower Keeps Rising” argues agents remove the coordination friction that kept a codebase’s shared language alive, so the tower keeps rising after understanding has collapsed.
  • Builder credibility: he experiments on Pi and CPython himself and publishes the tooling he actually uses.

Cautions
#

  • Cadence is essay-scale, not feed-scale; he will not tell you what shipped this week.
  • The flagship negative results come from single weekend experiments he designed himself, and he says so, but they get quoted as if they were benchmarks.
  • He is a vendor: Earendil makes Pi, so his harness comparisons carry an employer’s stake even when the criticism of other vendors is sharp.

Compared to
#

  • Simon Willison: Willison documents daily, Ronacher judges monthly; read Willison to know what changed, Ronacher to decide whether it mattered.
  • Steve Yegge: Yegge declares theses, Ronacher reports measurements and doubts; choose Ronacher when you want the case against as well as the case for.
  • Mario Zechner: Zechner builds the minimal core and argues for it in essays, Ronacher co-builds it and publishes the failure modes; read Zechner for the design, Ronacher for the stress test.

Bottom line
#

Recommended for engineers who want a measured, senior, skeptical counterweight to agent hype, grounded in published experiments. Not for daily tool churn, model news, or anyone who wants enthusiasm without invoices.

Top 5 recommended reading #

  • Astra for Coding: Why Are We Doing This Again? - the quantified skeptical case, publishing his own runtime hours and API spend where nothing of value emerged.
  • Better Models: Worse Tools - the deep dive that traces a tool-call regression down to in-band encoding and training inside a forgiving harness, then questions the tool schema contract.
  • The Tower Keeps Rising - the argument that agents remove the coordination friction that kept a codebase’s shared language alive.
  • The Coming Loop - his uneasy accounting of harness-level loops and what it means to ship code you cannot fully explain.
  • Before GitHub - the open-source history essay that shows the maintainer perspective underneath all his agent writing.

Changes
#

  • 2026-09-24 - Created after two earlier runs deferred him as a duplicate voice; the owner commissioned this note to pin the skeptical senior-engineer slot.
  • 2026-09-24 - Added the Top 5 recommended reading section.

See also
#

  • Simon Willison - the daily chronicler he acts as the monthly counterweight to
  • Mario Zechner - Pi’s author, whose project Ronacher co-builds and stress-tests
  • Thorsten Ball - the other practitioner-essayist comparison in this category
  • Pi - the harness he is the second-largest contributor to
  • Agentic Coding Tools Landscape - the tooling map his essays keep arguing with

References
#