---
session_id: "1323755"
title: "I deleted 90% of my harness (And It Got Better)"
phase: live
last_checked: 2026-11-12T13:24:21-08:00
render: 20261112-132421-c4d6ee
---

# Links: I deleted 90% of my harness (And It Got Better)

## Speaker

**Philipp Schmid**, Senior AI Developer Relations Engineer, Google DeepMind

- Site: [philschmid.de](https://www.philschmid.de/)
- X: [@_philschmid](https://x.com/_philschmid)
- GitHub: [philschmid](https://github.com/philschmid)

## Company

**[Google DeepMind (Gemini API)](https://ai.google.dev/gemini-api/docs)**: "Build
apps and agents with Gemini, Nano Banana, Gemini Omni, and Lyria using the Gemini
API documentation and reference." (from the docs; deepmind.google gives no title
or description to quote)

## Resources

- **Blog** · [How Foundational Models Became Superhuman in Bash](https://www.philschmid.de/superhuman-bash)
  (philschmid.de, 2026-09-01). His write-up of the same experiment: rebuild the
  harness, strip every tool but Bash and a media viewer, compare score, tokens and
  cost.
- **Blog** · [The Bitter Lesson](http://www.incompleteideas.net/IncIdeas/BitterLesson.html)
  (Rich Sutton). The essay the talk opens with: general methods that use
  computation win, so the harness should shrink as models improve.
- **Product** · [Gemini Decisions API](https://ai.google.dev/gemini-api/docs/decisions-api)
  (launched 2026-11-11, the morning of the talk). Typed yes/no, choice and score
  answers over text, images, audio, video and PDFs, with up to 1,000 questions per
  request.
- **Product** · [Gemini 4 Carbon](https://ai.google.dev/gemini-api/docs/models/gemini-4-carbon)
  (generally available 2026-11-10). The model in the bash-only demos. It chains
  2.2 to 3.7 commands per bash call, where Flash chains 1.3 to 1.8.
- **Repo** · [earendil-works/pi](https://github.com/earendil-works/pi). Pi, the
  minimal coding harness used for the demos (transcribed "PyHarness"). Its default
  read/write/edit/bash tools are what the talk compares against bash only.

## Gaps

- LinkedIn (`linkedin.com/in/philipp-schmid-a6a2bb196`, from the roster) answered
  999 with no page title naming him, so it is not corroborated and is left out.
- "DeepSVE" (the eval task source) and "Crossbench" were not identified.
- The QR code for free credits was not captured.
- No official recording or slides found yet. The dossier rechecks every 6 hours
  during the event.