Uploaded August 2026 | Updated September 2026, 3 weeks ago
Ben Kus stood on this same stage last year and recommended a graph based approach to building agents. Someone found him afterward to say it was exactly the answer he needed. Kus opens this talk by retracting it, and admits he has since wondered what became of that person. Nothing about the advice was wrong when he gave it. A better way simply arrived, as it did for most of what was presented at that conference. He runs the same exercise across models, agent design and retrieval, and each chain lands somewhere unrecognizable from where it started twelve months earlier.
The number that reframes it: infrastructure normally has a half life of three to five years, which is why the standard advice is to pick a stack, go deep and switch rarely, since every migration breaks something. Agent infrastructure has a half life measured in months. That breaks the usual advice, and it lands hardest on people. He tells it against himself, having asked an engineer to rebuild working agentic search in a new style, then proposed rebuilding it again the day after shipping. His answer is not a technology, it is preparation. Tell teams that change is normal rather than a failure, build abstractions that let the layer underneath be swapped, and only switch when eval sets say it matters, not when a paper is exciting. Box now reviews every AI technology on a six month clock.
Speaker info:
- https://x.com/benatbox
- linkedin.com/in/benkus
Timestamps:
0:00 - An exabyte, and now a trillion tokens
2:27 - The advice that used to work
4:48 - Retracting last year's talk from this stage
5:57 - How the model choices kept moving
7:07 - Agent design, from single calls to harnesses
8:15 - Retrieval, from keywords back to agents
10:31 - What does not change this fast
11:37 - Who a months long half life hurts
13:49 - Rebuilding it the day after shipping
14:57 - Telling teams that change is not a mistake
16:04 - Changing on evals, not on trends
17:13 - Judging vendors by how they handled change
Ben Kus stood on this same stage last year and recommended a graph based approach to building agents. Someone found him afterward to say it was exactly the answer he needed. Kus opens this talk by retracting it, and admits he has since wondered what became of that person. Nothing about the advice was wrong when he gave it. A better way simply arrived, as it did for most of what was presented at that conference. He runs the same exercise across models, agent design and retrieval, and each chain lands somewhere unrecognizable from where it started twelve months earlier.
The number that reframes it: infrastructure normally has a half life of three to five years, which is why the standard advice is to pick a stack, go deep and switch rarely, since every migration breaks something. Agent infrastructure has a half life measured in months. That breaks the usual advice, and it lands hardest on people. He tells it against himself, having asked an engineer to rebuild working agentic search in a new style, then proposed rebuilding it again the day after shipping. His answer is not a technology, it is preparation. Tell teams that change is normal rather than a failure, build abstractions that let the layer underneath be swapped, and only switch when eval sets say it matters, not when a paper is exciting. Box now reviews every AI technology on a six month clock.
Speaker info:
- https://x.com/benatbox
- linkedin.com/in/benkus
Timestamps:
0:00 - An exabyte, and now a trillion tokens
2:27 - The advice that used to work
4:48 - Retracting last year's talk from this stage
5:57 - How the model choices kept moving
7:07 - Agent design, from single calls to harnesses
8:15 - Retrieval, from keywords back to agents
10:31 - What does not change this fast
11:37 - Who a months long half life hurts
13:49 - Rebuilding it the day after shipping
14:57 - Telling teams that change is not a mistake
16:04 - Changing on evals, not on trends
17:13 - Judging vendors by how they handled change
