XGEN Newsletter vol.1

The basic unit of search moved from a list of links to a generated answer. To be found on the web now, you do not rank — you get cited inside an AI answer, and content has to be shaped for extraction and quotation rather than for clicks. A signal that our product documentation and engineering blog are due for a pass toward a form AI finds easy to cite.

The premise of the last three years — that AI is something you rent through an API — is wobbling. As frontier API pricing rises, companies that want predictable cost and their data kept inside move to open models plus their own infrastructure, which means the market for air-gapped, on-premise LLM platforms grows. Evidence that the direction we are building XGEN in matches where the industry is going — a quotable line for a proposal.

AI code audit has moved past PoC into a proven government-scale case. Cutting years of backlog to 20 hours reads less like replacing people and more like making a full sweep possible for the first time — something headcount had never allowed. The Claude Code backend that just landed in XGEN is for exactly this kind of work, which opens a new proposal scenario: auditing legacy code inside a customer's closed network.

Measured evidence that today's common design — the context window is big, so put everything in — may be wrong. Recall falling from 80% at 256K to 36% at 1M tokens means that as context grows the model becomes less able to *find* what it needs inside it. Worth keeping as a counterweight in the long-context era, and a reason our RAG principle of spending context sparingly through chunking and reranking still holds.

The assumption that combining models covers each other's weaknesses, refuted with data. Newer models are trained on similar data and therefore fail together on the same problems, so what matters in routing is not the second-best model but a model that fails differently. It is also why an LLM-as-judge should not use the same model family as the generator.

The unit of running an agent is moving from a single prompt to a loop with a stopping condition. Define the completion criteria up front instead of directing every step, and the agent repeats until it clears them. The hard problem shifts to how you define a good stopping condition — exactly the direction we are working in with the judgement gate in harness v2.