SATURDAY · AUGUST 29, 2026 · ISSUE 10 · YESTERDAY No. 10
Daily News — 2026-08-29
~7,700 packages released on PyPI·23 papers surfaced by Hugging Face
None of it is in your agent's weights.
Spotlight
A falsifiable argument that game engines are the missing verifier for spatial world models, backed by diagnostic pilots rather than definitive proof.
Spatial world models are stuck in an imitation loop. They train on scraped video, score outputs with CLIP or Fréchet Video Distance, and call that supervision.
JSX-as-timeline plus a media-inspection CLI is a genuinely coherent answer to the question of how agents should drive video editing.
The core bet here is that JSX is a better document format for video than a timeline GUI. A project is a folder of SolidJS components: a <stage> holds <scene> elements, clips and text carry timing as…
Papers
A falsifiable argument that game engines are the missing verifier for spatial world models, backed by diagnostic pilots rather than definitive proof.
A paper proposing game engines as executable verification and long-horizon trajectory sources for reinforcement learning post-training of spatial world models.
Penalizing disagreement is almost always right even when the pseudo-label is wrong — that asymmetry lets TTPO match label-supervised training without ever seeing a ground-truth answer.
TTPO is a label-free test-time training method that improves mathematical reasoning by distilling agreeing rollouts and penalizing disagreeing ones.
A rigorous survey that reframes agentic data quality around three asymmetric properties — validity first, then learner-relative difficulty, then behavioral coverage.
A framework for generating LLM agent training data by optimizing for accuracy, complexity, and diversity over raw scale.
A practical solution to a real deployment tension: training a compact agent to read its runtime configuration rather than memorize it, with production results that hold up.
TaoLive's Harness-Aware Training lets compact models track evolving digital-avatar harness configs with both low latency and high accuracy.
A joint video-and-action model for native game control that also names a concrete, previously undescribed failure mode in generative policies.
A unified model for video-game control that jointly predicts future visuals and executable keyboard-mouse actions using block-causal flow matching.
Separating oversight from execution mid-run, not just after it, is what makes the efficiency and recovery gains here actually compound.
PILOT lets a supervisor steer live agents mid-run and distill execution traces into reusable skills, enabling self-improvement without offline retraining.
9 more tool picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
Generating synthetic human videos from robot trajectories at scale is a more practical path to cross-task generalization than collecting paired demonstrations by hand.
A causal video-action model that uses in-context human video guidance to generalize robotic manipulation to unseen tasks.
ES preserves broad reasoning coverage where GRPO collapses, and the two methods compose better than either alone.
Evolution strategies beat GRPO on Pass@K and reasoning diversity through sparse functional updates and population diversity, with a hybrid approach outlined.
Persistent, never-rolled-back knowledge between raw traces and executable skills is what makes WikiSkill's gains both larger and more consistent than prior approaches.
WikiSkill co-evolves reusable agent skills with a persistent knowledge base so accumulated experience transfers across models.
Typed, solver-placed mates turn incremental LLM code generation into a 3D pipeline that out-edges every native generator tested.
A 3D modeling agent that generates editable, part-structured procedural assemblies with sharp geometry from text prompts.
Tools & packages
JSX-as-timeline plus a media-inspection CLI is a genuinely coherent answer to the question of how agents should drive video editing.
diffusionstudio/editor gives agents a programmatic interface to professional video editing.
A browser-automation bridge that makes a ChatGPT subscription do double duty as Codex's model backend — clever cost arbitrage with real fragility risk.
Routes ChatGPT Web (including Pro) as a native Codex model, unlocking context, tools, streaming, and images beyond Codex usage limits.
A genuinely full-stack AI classroom platform — the agent workbench and explicit tool surface make it more than a slide generator with voices.
Open-source multi-agent classroom that spins up an immersive, interactive learning environment in one click.
A coordination platform for multi-agent teams that keeps execution on the daemon and the control plane deliberately out of the hot path.
Tag-based agent coordination that lets teams invoke any agent inline, wherever work is already happening.
A native, zero-runtime code intelligence engine that turns structural graph queries into a genuine alternative to token-hungry file-by-file exploration.
Self-contained MCP binary that indexes codebases into a knowledge graph: 158 languages, sub-ms queries, and 99% fewer tokens with zero dependencies.
A tiered, versioned, self-auditing ruleset for stripping AI writing tells — more taxonomy than prompt, and honest about exactly what it cannot check.
A skill that audits and rewrites content to remove AI writing patterns, for use with agents like Claude Code and Codex.
10 more paper picks in this edition
Every pick in the Wire gets the same treatment: read, verified, and given a written verdict.
A structured agentic research layer for physics that enforces verification discipline and notation consistency across long multi-phase projects — already cited in published work.
Claims to be the first open-source agentic AI physicist, from Physical Superintelligence PBC (PSI).
A compiled-wiki alternative to RAG that bets on persistent, cross-linked knowledge over repeated retrieval — the tree indexing for long PDFs is the part worth watching.
OpenKB: an open-source knowledge base layer for LLM-powered applications.
Twenty markdown-based agent skills for job searching - the pattern is sound, the outcome claims are unverified, and the quality is opaque by design.
A collection of AI agent skills for resume writing, ATS optimization, interview prep, and job search, built for use with Claude Code.