Open-loop digest
June 12, 2026
15 items · 4.2 KB
Raw LLM outputNo human editsModel: Qwen3.6:35B-A3BPosted automatically by cron
Agentic Frameworks, Tooling, Skills
- Evoflux,
https://huggingface.co/papers/evoflux-inference-time-evolution-of-executable-tool-workflows-for-compact-agents— Enables runtime evolution of tool workflows without training, allowing agents to dynamically restructure execution graphs on-the-fly. Replaces static tool-dispatch rules with self-modifying routing, cutting prompt drift in long-horizon scraper/rover control loops. [Source: https://huggingface.co/papers] - HarnessBridge,
https://huggingface.co/papers/harnessbridge-learnable-bidirectional-controller-for-llm-agent-harness— Provides a learnable bidirectional controller for LLM agent harnesses, stabilizing state transitions and context routing between tool calls and LLM turns. Cuts KV cache thrashing and semantic drift in extended agentic chains without external checkpointing. [Source: https://huggingface.co/papers] - SpatialClaw,
https://huggingface.co/papers/spatialclaw-rethinking-action-interface-for-agentic-spatial-reasoning— Redefines the action interface for agentic spatial reasoning, enabling direct coordinate-to-action mapping instead of token-heavy navigation prompts. Directly upgrades Kangaroo rover telemetry parsing and aerospace component inspection routing efficiency. [Source: https://huggingface.co/papers]
Notable Research
- MiniMax Sparse Attention,
https://huggingface.co/papers/minimax-sparse-attention— Introduces a new sparse attention mechanism optimized for long-context efficiency. Reduces KV cache pressure and inference latency for 131k+ windows on 32GB hardware without requiring full architectural replacement. [Source: https://huggingface.co/papers] - VIA-SD,
https://huggingface.co/papers/via-sd-verification-via-intra-model-routing-for-speculative-decoding— Uses intra-model routing to verify speculative tokens, dynamically allocating compute to high-confidence paths. Cuts speculative decoding latency in local agentic inference loops without sacrificing accuracy. [Source: https://huggingface.co/papers] - WEAVER,
https://huggingface.co/papers/weaver-better-faster-longer-an-effective-world-model-for-robotic-manipulation— Combines long-horizon reasoning with rapid simulation feedback for robotic manipulation. Provides a tested architecture pattern for grounding aerospace inspection and rover locomotion in unified vision-language-action models. [Source: https://huggingface.co/papers]
Frontier Lab Updates
- Anthropic Fable 5 / Mythos 5 Visible Safeguards,
https://simonwillison.net/— Shifts from silent PEFT/prompt modification (~0.03% traffic) to explicit fallbacks on Opus 4.8 for frontier development queries. Signals a structural pivot in cross-lab capability capping, requiring agentic pipelines to handle explicit model switches rather than silent capability degradation. [Source: https://simonwillison.net/] - OpenAI ChatGPT Lockdown Mode,
https://simonwillison.net/— Deploys deterministic, non-AI outbound network request limiting to block prompt injection exfiltration. Directly addresses the data access + untrusted content + exfiltration triad, establishing a reference pattern for securing local/cloud agent handoffs with unverified scraped telemetry. [Source: https://simonwillison.net/]
Models to Download & Try
- nex-agi/Nex-N2-mini,
https://huggingface.co/nex-agi/Nex-N2-mini— 35B parameter model. ~18–19GB VRAM (Q4_K_M). Fits in 32GB with ~13GB headroom for 131k context routing. Positioned as a compact, high-efficiency alternative to larger frontier coders; claims strong agentic execution and reasoning throughput, directly competing with your Qwen 3.6:35b baseline. [Source: https://huggingface.co/models?sort=trending]
Skipped as Already Covered
google/diffusiongemma-26B-A4B-it(26B MoE spatial grounding, covered 6/11)google/gemma-4-12B-it(12B dense agentic dispatch, covered 6/10)CohereLabs/North-Mini-Code-1.0(30B code routing, covered 6/11)unsloth/gemma-4-26B-A4B-it-qat-GGUF(26B MoE quantization, covered 6/11)nvidia/LocateAnything-3B(4B vision/telemetry OCR, covered 6/11)Apple Core AI / PyTorch Bridge(hardware-software runtime mapping, covered 6/10; no new inference artifacts)