Andrej Karpathy’s Math Proves Agent Skills Will Fail. Here’s What to Build Instead.

Uploaded: 2026-03-21
The video discusses the future potential of AI in executing complex business workflows by achieving high reliability and dependability. It emphasizes the concept of agent skills and harness engineering to improve the effectiveness of AI systems in tasks like compliance audits and contract reviews, ultimately showing how structured approaches can mitigate failures and enhance performance.

Add to wishlistAdded to wishlistRemoved from wishlist 0
Add your review

A technical walkthrough of creating reliable, scalable AI harnesses to execute complex, multi-stage workflows (e.g., contract review), covering architecture, delegation, validation, and cost/control trade-offs.

– Demo of a contract-review harness with sub-agents, virtual file system, and state management.
– Harness patterns: deterministic rails, parallel sub-agents, memory and context strategies.
– Focus on reliability metrics, validation loops, model selection, and cost optimisation.

Quotes:

The march of nines: each extra nine takes comparable engineering effort.

Agentic workflows compound failure: a 10-step flow at 90% per step yields ~6 daily failures.

Put AI on deterministic rails — Stripe merged 1,300 pull requests weekly by validating changes against tests.

Statistics

Upload date:2026-03-21
Likes:2370
Comments:117
Fan Rate:4.27%
Statistics updated:2026-04-16

Specification: Andrej Karpathy’s Math Proves Agent Skills Will Fail. Here’s What to Build Instead.

channel

top

Andrej Karpathy’s Math Proves Agent Skills Will Fail. Here’s What to Build Instead.
Andrej Karpathy’s Math Proves Agent Skills Will Fail. Here’s What to Build Instead.