Five things I bookmarked this week. Each one covers a different angle — agent frameworks, frontier models, and tooling — so there's some breadth here.
1. NVIDIA NOOA — agents as plain Python classes
NVIDIA Labs released NOOA (NVIDIA Object-Oriented Agents) on August 7 under Apache 2.0. The core design decision: an agent is a single Python class, not a YAML config or a JSON scaffold. Six harness capabilities — typed I/O, pass-by-reference, code as action, programmable loops, explicit state, and model-callable APIs — drive the design. LiteLLM handles model routing, so the same agent code runs against Claude, GPT-5, Ollama, or a local vLLM endpoint without changes. Install is pip install nooa (v0.0.8, alpha, Python 3.12–3.13).
The number they lead with — 82.2% on SWE-bench Verified using GPT-5.5, with half the token budget of prior state-of-the-art — is worth paying attention to. I'm more interested in whether the object-oriented model holds up outside coding tasks, where SWE-bench stops being useful signal. That's the question I'll be watching through the rest of the year. Source: MarkTechPost.
2. PrimeIntellect/prime-agent — self-improving RLM harness
Prime Agent dropped August 5, MIT license. The architecture is unusual enough to read past the headline. Rather than a tool-call loop, the agent runs inside a persistent IPython kernel. Context is a variable the agent actively manages; memory, skill definitions, and subagent specs are durable state it can update (CRUD) from within its own trajectory. They call this the Recursive Language Model (RLM) abstraction.
With Claude Opus 5, they report 95.5% on ARC-AGI-3, fractionally above the 95.4% human expert baseline. I'm generally cautious about single-benchmark numbers, but the architectural difference from ReAct-style agents is real enough that the technical report is worth reading rather than skimming. The repo is early-stage; the GitHub star count has been climbing fast since the August 5 announcement.
3. Qwen3.8-Max — 2.4T parameter MoE from Alibaba
Alibaba published Qwen3.8-Max on August 3 (MarkTechPost). Architecture is MoE: 2.4 trillion parameters total, 95 billion active per token. Context window is 1 million tokens. Native multimodal support is included.
Alibaba is pricing international API access at roughly 40% of Claude Opus 5 input and 24% output — aggressive positioning. Benchmark positions on launch: fifth Text Arena, second Vision Arena. Open weights are scheduled for release this week; I'll form a more concrete view once I can run it locally rather than through the API. The pricing point means it will probably land in my comparison ETL on aiappdex.com — I've been adding new frontier models to the pairwise compare pages as weights become available.
4. DeepSeek V4-Flash 0731 — official GA with re-post-trained agent checkpoint
DeepSeek moved V4-Flash from preview to official general availability on July 31 (MarkTechPost). Same 284B CSA+HCA backbone as the preview checkpoint; what changed is a re-post-training pass focused on agentic tasks. DeepSeek reports official V4-Flash beating V4-Pro-Preview across nine agent benchmarks. The model is MIT-licensed on Hugging Face and the API is at $0.14/$0.28 per 1M input/output tokens.
I switched my comparison ETL to the preview three weeks ago and the official checkpoint runs noticeably cleaner on instruction-following edge cases — specifically the cases where the preview would produce partial JSON before hitting a stop sequence. That single improvement is enough to keep it in my pipeline for the foreseeable future. The price point also makes it viable for high-volume extraction tasks where using a flagship model would be economically absurd.
5. addyosmani/agent-skills — 76k GitHub stars, production skills for coding agents
agent-skills by Addy Osmani (Engineering Lead at Google) crossed 76k GitHub stars this week and has been one of the fastest-growing developer tool repos of 2026. Twenty-four skills in plain Markdown — security review, migration strategy, testing discipline — compatible with Claude Code, Cursor, GitHub Copilot, and Codex. Install: npx skills add addyosmani/agent-skills.
What distinguishes it from other prompt template repos: the skills are specific about what to check, not vague about intent. Rather than "be thorough," the security-review skill names concrete vulnerability classes. I've been incorporating a few of the patterns into how I prompt Claude Code in my own pipeline. Worth reading even if you don't use the installer — the framing is useful standalone.
Part of an ongoing 6-month experiment running three AI-curated directory sites. The technical claims here are real; this article was AI-assisted.
Top comments (0)