A clearer view of AI.
THE INDEPENDENT AI.

WORTH TESTING · 28 OCTOBER 2026

Andrej Karpathy at Sequoia's AI Ascent: how to work with agents

A genuine 30-minute stage interview with Andrej Karpathy at Sequoia Capital's AI Ascent, reposted by @hanakoxbt. The repost's framing is false: Karpathy never mentions 'Opus 5.5' (his one example is 'Opus 4.7') and never says anything about being '99% automated'. Verdict: excellent mental models (Software 3.0, verifiability, specs over vibes, agents as interns) but a talk, not a tutorial; no skills needed to follow it, real software skills to apply the agentic-engineering parts. The timestamps below are approximate.

Try this

  1. Give the agent bigger chunks of work instead of reviewing line by line. 01:45
  2. Treat the context and prompt as the program you're writing. 03:45
  3. Paste raw install instructions into an agent instead of writing setup scripts. 04:28
  4. Feed raw inputs (such as a photo) straight to a multimodal model for the final output. 05:58
  5. Let an AI turn a pile of raw documents into a company wiki. 07:16
  6. Aim agent workflows at tasks whose results can be checked. 10:30
  7. AI ability is 'jagged': keep a verification step in the loop. 12:30
  8. Map what your agent can and can't do by testing it on your own tasks. 13:30
  9. Agentic engineering: write detailed specs and review the output. 16:34
  10. Manage the agent like an intern and keep the taste and judgment calls yourself. 20:07

Timestamps point to the moment in the source video.

Cost

Free to watch.

Source

Hype or Fact?

Hype Karpathy at AI Ascent: 'Opus 5.5' and '99% automated'

How we knowSteps to try, not a claim that we independently tested every result. Follow the linked sources and check current prices. Our methods & sources ↗