AI Briefing — 2026-08-14

🔬 Research

Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill

271 upvotes · Zhuoyang Qian et al. · Paper

A system that turns a research idea into a full manuscript — including literature retrieval, experiment design & execution, figure generation, and claim revision — implemented as 13 composable skills inside an existing coding assistant, with no separate agent platform needed. Key reliability features include deterministic integrity checks and a mechanism to cap the "Self-Refutation Loop" where repeated experiments keep rejecting the original hypothesis. Achieves 99.5% citation validity, bumps fabrication detection from 14% (single-pass draft) to 92% (full review stack), and costs ~$8.1 and 3.2 hours per paper at 11.9M tokens. The cost and composability make it a credible baseline for automated research pipelines rather than a research curiosity.

OpenART: Scaling Agent Red Teaming via Open-Ended Environment Evolution

250 upvotes · Yunhao Chen et al. · Paper

OpenART is a red-teaming benchmark for AI agents that emphasizes stateful, long-horizon tasks — scenarios where early actions poison future decisions — rather than the short static prompts most safety benchmarks use. It ships 10,000+ validated scenarios across 50 domains, drawing from 500,000+ tools/skills, with a median task requiring 97 tool calls across 75 agent-model configurations. The proposed black-box attack policy (EMHA) reaches an 85% Attack Success Rate overall, with its advantage over instruction-only attacks growing from ~2% on simple tasks to 17%+ on complex ones. Notably, the specific runtime implementation of an agent explains more safety variance than the underlying model — a finding with direct implications for how teams deploy agents in production.

AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

105 upvotes · Cheng Qian et al. · Paper

The paper asks whether a stronger model can improve a weaker model's performance without any fine-tuning, purely by constructing inference-time "harnesses" (structured prompts, deterministic code, routing logic, strict output formats). Tested on four Theory-of-Mind benchmarks, this approach nearly doubles weaker-model accuracy from 0.49 to 0.91 on average, with the largest gains going to the weakest targets. The gains come mostly from offloading unstable reasoning into deterministic code and strict formatting — not from chain-of-thought or broader sampling. This is a practical alternative to distillation for teams that cannot retrain models, and scales with the builder model's reasoning budget.

🛠 Tools

cathrynlavery/diagram-design — Editorial SVG diagram types for AI coding agents

4,475 stars/day · ⭐ 15,795 total · Repo

A collection of 27–29 opinionated diagram types (architecture, loops, flowcharts, pyramids, trust boundaries…) delivered as self-contained HTML/SVG files and a single skill file for Claude Code, Codex, and similar agents. It reads your website to match your brand in ~60 seconds, supports redrawing existing draw.io or Mermaid diagrams, and avoids generic rounded-box aesthetics. Version 2.0 added "the Loop" pattern (flywheels with shared-memory hub), and v2.3 added semantic system patterns. A companion native app (macOS/Linux/Windows) lets you browse and install agents into your tool of choice with one click.

macro-inc/macro — Unified AI workspace (email, chat, docs, tasks, CRM, calls)

1,239 stars/day · Repo

Macro aims to replace the patchwork of SaaS tools teams use by combining email, chat, documents, tasks, agent workflows, calls, and CRM in one Rust-built workspace with shared AI memory and @-linking between entities. The description is high-level, but the Rust backend and multi-modal scope position it as a performance-focused alternative to tools like Notion + Linear + Superhuman. Worth watching if you're evaluating AI-native productivity stacks for engineering teams.

msitarzewski/agency-agents — Curated library of specialist AI agent personas

778 stars/day · ⭐ 145,351 total · Repo

A large and actively maintained collection of AI agent "personalities" (frontend engineers, community managers, reality checkers, etc.), each defined with a specific domain, communication style, and deliverable-focused workflow. Agents install directly into Claude Code, Cursor, Codex, Gemini, and others via a native cross-platform app. Born from a Reddit thread, it has grown to 145K stars, indicating strong community adoption. MIT-licensed; useful as a starting point for teams building specialized agent pipelines without crafting system prompts from scratch.

📡 Analysis

sqlite-utils 4.2 (+ 4.2.1 hotfix)

2026-08-13 · Post · tags: releases, sqlite, sqlite-utils

sqlite-utils 4.2 ships significant improvements to table.transform() — the feature that handles complex ALTER TABLE operations by rebuilding tables — now preserving check constraints, unique constraints, and column comments across transforms. New introspection properties for check constraints are also added, alongside smaller fixes from five community contributors. A crashing bug introduced in 4.2 was patched immediately in 4.2.1, so users should go straight to 4.2.1.

llm-gemini 0.33

2026-08-13 · Post · tags: google, gemini, llm, generative-ai

Simon Willison's llm-gemini plugin now supports the newly released Gemini 3.7 Flash, plus gemini-3.6-flash, gemini-3.5-flash-lite, and two embedding models (gemini-embedding-2, gemini-embedding-001). The update also bumps compatibility to LLM 0.32, unlocking reasoning trace visibility and server-side tools (e.g., CodeExecution) directly from the CLI. A notable caveat: Gemini 3.7 Flash drops the minimal thinking-effort tier that 3.6 Flash had, leaving only medium and high effort modes.

Sources: HuggingFace Papers API, GitHub Trending, simonwillison.net