How Will We Prompt AGI? A History of Harness Hacks
summary
A concise history of harness hacks traces concrete wins and tradeoffs: Viktor runs a single continuous agent with 173 integrations, 5,400 tools, 1,300 skills automating half a company; GPT‑3 few‑shot and prompt programming lead to chains, ReAct loops and Tree of Thoughts which lift GPT‑4 from 4% to 74% on game of 24 at higher cost; constrained decoding and grammar parsers enforce valid JSON; models optimize prompts and design harnesses reaching 77.4% on Sweetbench and speculative programmatic tool calling yields ~1.2x–2x speedups. These results push work toward graph structures, hill‑climb gradients, and swarm scale.
AI-written summary · about
chapters (24)
- 0:00Prompting the audience
- 1:00Viktor's continuous agent
- 2:40What counts as a harness?
- 3:12GPT-3 and few-shot examples
- 4:14Prompt programming
- 4:53AI chains
- 5:32Chain-of-thought prompting
- 6:10ReAct and agent loops
- 6:52Tree of Thoughts
- 7:45Constrained JSON generation
- 8:53Models optimizing their own prompts
- 9:46Best-of-N generation
- 10:31Splitting work across context windows
- 11:17Letting models modify the harness
- 12:32Speculative code execution
- 13:48Three trajectories of harness evolution
- 14:00From prompts to graphs
- 14:45The model absorbs the harness
- 15:40Scaling from one model call to swarms
- 16:05Turning tasks into hill climbs
- 17:06Work becomes a graph
- 17:50The world retreats into the model
- 18:43A civilization of experts
- 19:22How we may prompt AGI
related talks
47mThe Art of Prompt Engineering: Best Practices for Effective AI Interactions
22mAI, Make It So - A Dive into the Universe of Prompt Engineering
01h 08mIn Prompts We Trust
48mStop Fighting Your AI Engineering Prompts That Actually Work
02h 01mPrompt Engineering and AI Red Teaming
24mBetter Prompting with Pseudocode
20mThe Prompt Is Still a Punch Card
44mThe Science of Effective Prompting