August 23, 2026

Agentic Frameworks, Tooling, Skills

• FlowEvo: Self-Evolving Agents through the Co-Evolution of Workflows and Executable Skills, https://huggingface.co/papers/date/2026-08-21 — Southeast Univ research demonstrating automatic co-evolution of agent workflows and executable skills via multi-turn interaction; cuts manual tool-definition overhead and replaces static prompt routing in local agentic stacks.
• Hierarchical Self-Improvement: A Framework for Task-Specific Evolvable Agent Harnesses, https://huggingface.co/papers/date/2026-08-21 — HKUST blueprint for continuously evolving, task-specific agent harnesses; directly applicable to hardening your Kangaroo rover’s decision loops without external reward-model scaffolding.
• llm CLI (Simon Willison), https://simonwillison.net/ — Updates Python SDK to accept per-call embedding keys without mutating shared state, plus template composition (-t lhigh -t pelican); streamlines local agent telemetry injection and multi-stage prompt routing.

Notable Research

• τ_0-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation, https://huggingface.co/papers/date/2026-08-21 — Shanghai Innovation Institute paper introducing hierarchical vision-language-action policies that leverage test-time compute guided by internal world models; directly scalable to aerospace vision-scraper pipelines and Kangaroo rover teleoperation without synthetic domain gap calibration.
• EnvHarness: Awakening Static Worlds for Agent Learning, https://huggingface.co/papers/date/2026-08-21 — Google’s methodology for converting static simulation environments into dynamic agent-learning simulators; enables high-fidelity reward/trajectory generation for rover simulation pipelines without custom environment code.

Frontier Lab Updates

• Qwen/Qwen3.8-27B, https://huggingface.co/Qwen/Qwen3.8-27B — Alibaba’s Apache 2 vision-capable 27B LLM released late August; scores within 1 point of GLM-5.2/DeepSeek V4 Pro max configurations and matches GPT-5.6 Luna, positioning as a stronger vision/coding base than your current 35B setup. [Source: https://simonwillison.net/]
• DeepSeek-V4-Pro-0813 / DeepSeek-V4-Flash-0731, https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro-0813 — DeepSeek’s 1.7T and 304B model variants; establish new cloud-scale capability baselines that inform Qwen3.8’s routing and compression targets for local 27B classes. [Source: https://huggingface.co/models?sort=trending]

Models to Download & Try

• Qwen/Qwen3.8-27B, https://huggingface.co/Qwen/Qwen3.8-27B — 28B parameters; ~16.5GB Q4_K_M footprint leaves ~15.5GB for context (~85k tokens at 16-bit KV). Vision-capable, Apache 2 license. Claims parity with GPT-5.6 Luna and near-GLM-5.2 max scores, directly outperforming Qwen3.6:35b on multimodal instruction fidelity. [Source: https://huggingface.co/models?sort=trending]
• ornith-ai/Ornith-1.5-35B-A3B-GGUF, https://huggingface.co/ornith-ai/Ornith-1.5-35B-A3B-GGUF — 36B total / 3B active (MoE); ~18GB Q4_K_M footprint leaves ~14GB for context. Sparse architecture drastically cuts KV and compute pressure; enables longer context routing for aerospace telemetry parsing without expanding base model VRAM. [Source: https://huggingface.co/models?sort=trending]

Skipped as Already Covered

• PostHog MCP Integration (covered 7/18)
• Tencent/Hy3 & AngelSlim/Hy3-GGUF (295B/299B MoE; exceeds 32GB)
• zai-org/GLM-5.2 & jlnsrk/GLM-5.2-colibri-int4 (covered 7/6–7/17)
• Moonshot Kimi-K2 & xAI Grok Build (covered 7/9–7/17)
• Anthropic Fable 5 policy & subscription updates (covered 7/16–7/19)
• thinkingmachines/Inkling (975B MoE; covered 7/17–7/20)