The paperwork is being rewritten, for agents, not humans

Markup for machines

Every prompt is currently a soup: instructions, data, context, and user text blended into one undifferentiated string. LLMON, a preprint proposing an LLM-native markup language, argues the interface should grow up — separating instructions from data in LLM prompts, the way HTML separated structure from presentation. The authors claim three wins: a cleaner attack surface, since structured markup can mitigate prompt injection; better accuracy, from semantic metadata available during training and inference; and a contract both sides of the interface can rely on. It is a small idea with large implications: prompts become documents with a schema instead of vibes in a text box. m:…4617

The security angle is the wedge that could make it stick. Prompt injection works because the model cannot tell instruction from data; a markup language that separates the two turns a social-engineering problem into a parsing problem. Structured interfaces do not just make agents faster — the SPRA token cuts show that — they make the boundary between command and content enforceable. Expect the first production LLMON-like interface to be sold on security, not on accuracy. m:…4617

Papers, recompiled

The scientific paper is the next format up for recompilation. SPRA — Structured Paper Representation for Agents — asks whether agents need a new kind of paper, and answers by organizing over a million AI papers into 19.95 million research units and 24.07 million internal relations. The payoff is measurable: source-conditioned scientific QA improves by 0.381 out of 5 over passage retrieval, while input-plus-output tokens drop 28.5 percent versus passage retrieval and 95.3 percent versus full-document compaction in cross-paper reasoning. The PDF was designed for eyes. SPRA is designed for attention heads. m:…5452

The same recompilation is hitting 3D worlds. 3DCodeVerse treats 3D scenes as code: fine-tuning on its 2M-example dataset significantly improves the 3D coding capabilities of open-source models, and its harness — which evaluates and revises generated programs automatically using execution checks, spatial measurements, and visual feedback, with no human in the loop — wins 90 percent of pairwise human evaluations against Claude Design on 3D objects. Worlds you can diff, test, and revise: the asset pipeline becomes a software pipeline. m:…8735

The part agents still fumble

The sobering entry in the pile is Neurodata Without Boredom, a benchmark of agentic AI on data reuse. Coding agents succeed on most subtasks but rarely produce a fully error-free end-to-end solution for data reformatting. Agent judges are unreliable at catching errors on their own, though they can automate process-based evaluation when given reference solutions. And aggregate metrics fail to capture scientifically significant errors — the scoreboard says fine while the science is wrong. It is a useful warning for the whole agent-native-format movement: structured inputs make agents faster, but the last mile of correctness in data work is still unclaimed territory. m:…4508

The benchmark also indicts the judges. Agent judges are unreliable at catching errors on their own, the authors found — but they can automate process-based evaluation when given reference solutions. The practical read: do not ask an agent whether the data work is right. Give it the reference, ask it to check the process step by step, and keep a human in the loop for the scientifically significant calls that aggregate metrics miss. m:…4508

The decision for builders is straightforward: publish for the reader you actually have. If agents are the primary consumers of your docs, data, and papers, ship structured, machine-checkable formats — and budget for verification, because the agents will happily reformat your data into something confidently, subtly wrong. m:…4617 m:…5452