AI Briefing — 2026-07-17

🔬 Research

Harness Handbook: Making Evolving Agent Harnesses Readable, Navigable, and Editable

177 upvotes · Ruhan Wang et al. · Paper

As AI agents grow more complex, maintaining and evolving their "harness" code (prompt construction, state management, tool invocation) becomes a major bottleneck — developers struggle to map high-level behaviors to concrete code locations. The paper introduces the Harness Handbook, a behavior-centric representation generated via static analysis and LLM-assisted structuring, paired with Behavior-Guided Progressive Disclosure (BGPD) to navigate from intent to implementation. This improves localization accuracy and edit quality while consuming fewer planner tokens, with the biggest gains on scattered, cross-module, or rarely executed code paths. A timely contribution as production agentic systems grow harder to maintain.

Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

116 upvotes · Guoxuan Chen et al. · Paper

Boogu-Image-0.1 is an open-source family of unified multimodal models (Base, Turbo, Edit, Edit-Turbo) covering text-to-image generation, fast inference, instruction-based editing, and bilingual (Chinese-English) rendering. Notably, it was trained on only ~208M unique images at an estimated base cost of ~$400K, yet matches or exceeds other open-source models and approaches leading closed-source systems like GPT-Image-2. Weights, code, and training recipes are released under Apache 2.0, making it a strong open alternative to proprietary pipelines. The paper also shares practical insights on data quality, training pipelines, and agentic inference-time scaling under tight compute budgets.

Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models

93 upvotes · Yubo Wang et al. · Paper

Standard left-to-right pretraining doesn't naturally teach models to integrate external tool return values mid-reasoning — a core skill for coding agents. This work exploits the structural analogy between an agent's action-observation-continuation loop and a function call site, using a function-aware Fill-in-the-Middle (FIM) mid-training objective on a 2.6B-token decontaminated corpus from 968 GitHub repos. Applied to Qwen2.5-Coder-Instruct (7B/14B) and Qwen3-8B, it yields +2.8 to +5.4 points on SWE-Bench (Verified and Lite) across multiple post-training pipelines, while also reducing the capability erosion that agentic fine-tuning usually causes on general coding benchmarks.

🛠 Tools

Nutlope/hallmark — Anti-AI-slop design skill for coding agents

3,372 stars/day · Repo

Made by Together AI, Hallmark is a design skill (a shareable prompt/rule-set) for Claude Code, Cursor, and Codex that actively resists generating generic-looking AI UI. It selects from 20 themes and multiple macrostructures, runs 57 "slop-test gates" plus a self-critique pass before emitting code. Four verbs cover building new UI, auditing existing code, redesigning with a different fingerprint, and extracting DNA from a design screenshot. At 11,469 total stars and 3,372 gained in a single day, this is clearly resonating with developers tired of cookie-cutter AI-generated frontends.

mattpocock/skills — Composable Claude agent skills for real engineering workflows

2,060 stars/day · Repo

Matt Pocock (Total TypeScript) has open-sourced the agent skills he uses daily from his .claude directory, installable in 30 seconds via npx skills@latest add mattpocock/skills. Unlike heavier frameworks (GSD, BMAD, Spec-Kit), these are intentionally small, composable, and model-agnostic, designed to keep engineers in control rather than delegating the full process to an AI. With ~175K total stars and 60K newsletter subscribers, this repo has become a de facto reference for production-grade agentic developer workflows.

Graphify-Labs/graphify — Turn any codebase into a queryable knowledge graph for AI agents

1,107 stars/day · Repo

Graphify is a coding-assistant skill compatible with Claude Code, Cursor, Gemini CLI, Codex, and others that ingests code, SQL schemas, shell scripts, docs, images, or videos and builds a unified, queryable knowledge graph across all of them. It allows AI coding agents to reason about app code, database schema, and infrastructure together in a single graph — addressing a key limitation of file-based context retrieval. The project sits at ~89K total stars and is actively maintained (last push yesterday), signaling strong community traction.

📡 Analysis

Kimi K3, and what we can still learn from the pelican benchmark

16 juillet 2026 · Post · tags: ai, llms, llm-pricing, moonshot, kimi, ai-in-china

Moonshot AI released Kimi K3, a 2.8 trillion-parameter MoE model they claim is the first "open 3T-class model," beating DeepSeek v4 Pro in raw size and scoring just below Claude Fable 5 on long-horizon knowledge work (Elo 1547 on Artificial Analysis). At $3/M input · $15/M output, it's the priciest model from a Chinese lab to date — a 15× price increase over Kimi K2.6 — and an open-weight release is promised by July 27. Simon also benchmarks it with his signature "pelican riding a bicycle" SVG test: K3 produced a competent result but burned 13K reasoning tokens, costing 25 cents for a single SVG.

Firefox in WebAssembly

16 juillet 2026 · Post · tags: browsers, firefox, webassembly, ai-assisted-programming, claude

Puter compiled Firefox/Gecko entirely to WebAssembly so a full browser runs inside another browser, using the Wisp WebSocket protocol to proxy network traffic through Puter's servers (required since browser-sandboxed code can't open arbitrary connections). The project consumed an estimated $25,000 worth of Claude Opus and Fable tokens (much less in actual dollars thanks to a Claude Max subscription), illustrating both the power and the cost structure of AI-assisted systems programming. End-to-end encryption for HTTPS traffic is preserved through the proxy, which is a non-trivial achievement.

Spot birds not golf

17 juillet 2026 · Post · tags: ai-energy-usage, ai

Simon does quick back-of-the-envelope math on hyperscaler water consumption: Google used 10.9 billion gallons in 2025 (~30M gallons/day), while the 120 golf courses of the Coachella Valley each consume ~750K gallons/day — meaning buying out ~40 of them would roughly offset Google's usage. It's a tongue-in-cheek post, but the underlying numbers are real and serve as a concrete reminder of the physical resource footprint of large-scale AI infrastructure — a topic regulators and investors are increasingly scrutinizing.

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