Armin Ronacher of Earendil uses his own coding-agent sessions to argue for a better software workshop. This 20-minute AI Engineer talk connects human attention, engineering judgment, shared formats, and recoverable automation.

Better models, the same core tools

  • Ronacher says Pi’s core tool set stayed stable while improved models needed fewer corrections and leaned more heavily on Bash.
  • His sessions’ 90th-percentile agent run time grew from about 2½ minutes to 14 minutes or more. This measures duration, not task quality.
  • The claim concerns the core tools; Pi also gained orchestration features such as code mode, discussed later in the talk.

Starting projects is easier than caring for them

  • Ronacher reports roughly 120–190 prompts a day even as total agent activity grew beyond his working hours.
  • He started 141 repositories, but regularly returned to only about 10–15. His informal conversations suggest others face a similar limit.
  • A friend’s automated open-source maintenance loop was switched off when interest faded and its token budget became more valuable elsewhere.
  • More agents also put pressure on code hosts and older websites that were built for lower request volumes.

Keep engineering judgment in the loop

  • Push back when a completed task adds complexity or produces features that fail to fit together.
  • Ronacher compares uncontrolled feature accumulation to the Simpsons’ overloaded car; matching individual requests can still produce an incoherent whole.
  • He asks agents for short executive summaries and readable HTML explanations before verification, around 70% through a run.
  • Those explanations let him follow work while it finishes. Their accuracy still needs checking against the implementation.

Preserve the ability to work together

  • Cheap custom implementations can fragment shared formats; Ronacher cites a tar archive interpreted differently by different implementations.
  • Open protocols, compatible files, and upstream collaboration require agreement between people even when implementation becomes cheap.
  • Aim the extra capacity at reliable, fast, memory-efficient software. Ronacher sees promise in verification and memory-safe languages while acknowledging immature tooling.

Make long-running work recoverable

  • Agents working overnight need to restart, suspend, and resume without constant human attention.
  • Flexible inference scheduling can introduce long waits; suspending the sandbox too can avoid paying for idle resources.
  • Ronacher proposes this as infrastructure worth improving. He supplies no recovery benchmark or pricing comparison here.

Let programs handle repeated decisions

  • Code mode lets an agent write JavaScript to compose tool calls and filter results before they enter the model’s context.
  • His example routes 3,000 support requests through a generated workflow, using Jev classifiers for bounded decisions and agents for cases needing more work.
  • Pi’s documentation confirms that scripts can invoke tools and classifiers inside a restricted JavaScript sandbox.
  • Caveat: the session numbers are Ronacher’s personal observations. The support-routing example has no measured accuracy or cost result in this talk.

“Now that we can build everything, maybe we should not.” — Armin Ronacher, 19:23