vaniagarwal.dev

Vani
Agarwal

i design, build, and evaluate AI systems

Now showing
Harness-Bench results for Claude Fable 5.2 in four setups. Rubric score: closed-book 0.558, plain tools 0.518, GPT Researcher 0.433, Deep Agents 0.562. Mean minutes per run: 2.0, 5.1, 7.4 and 12.8. Mean cost per run: $0.45, $2.93, $7.80 and $8.02.Harness-Bench results for Claude Fable 5.2 in four setups. Rubric score: closed-book 0.558, plain tools 0.518, GPT Researcher 0.433, Deep Agents 0.562. Mean minutes per run: 2.0, 5.1, 7.4 and 12.8. Mean cost per run: $0.45, $2.93, $7.80 and $8.02.
Harness-Bench on research.siteBenchmark, 4 harnesses, 12 questions · Nov 2026
Harness-Bench
currently atINTDEV, Internet Development StudioINTDEV, Internet Development Studio

Past work

research.site

Writing

Why do benchmarks remain the industry’s public standard?Essay, 16 Nov 2026
  • The Fountainhead, Ayn Rand
  • 84, Charing Cross Road, Helene Hanff
  • The Forty Rules of Love, Elif Shafak
  • Steve Jobs, Walter Isaacson
  • Bottle of Lies, Katherine Eban
  • 21 Lessons for the 21st Century, Yuval Noah Harari
  • Blue Ocean Strategy, W. Chan Kim and Renée Mauborgne
  • Kafka on the Shore, Haruki Murakami
  • The Diary of a Young Girl, Anne Frank
  • My Life in Full, Indra Nooyi
  • Becoming, Michelle Obama
  • Bad Blood, John Carreyrou
  • The Romantic Manifesto, Ayn Rand
  • 1984, George Orwell
  • The Myth of Sisyphus, Albert Camus
  • Thus Spoke Zarathustra, Friedrich Nietzsche
  • Norwegian Wood, Haruki Murakami
  • The Visual Display of Quantitative Information, Edward Tufte
  • Competitive Programming 3, Steven Halim and Felix Halim
  • Introduction to Algorithms, Cormen, Leiserson, Rivest and Stein
  • The Non-Designer’s Design Book, Robin Williams
  • Thinking with Type, Ellen Lupton
  • Arts of Living on a Damaged Planet, Anna Tsing, Heather Swanson, Elaine Gan and Nils Bubandt
  • The Unsettling of America, Wendell Berry

Contact