Latent Space AI News：Anthropic

if you see this, it’s beacuse you’re a real fan.

AI News for 9/16/2026-9/17/2026. We checked 12 subreddits, 544 Twitters and no further Discords. AINews’ website lets you search all past issues. As a reminder, AINews is now a section of Latent Space. You can opt in/out of email frequencies!

AI Twitter Recap

Agent Runtimes, Long-Horizon Workflows, and the Rise of Coordinator UIs

• Claude Code Projects pushes “one conversation, many cloud threads” into product: Anthropic rolled out Projects in Claude Code, where a single conversation can spawn parallel cloud sessions, pass context between threads, and continue running after the user leaves. Follow-up posts clarify availability and that threads currently run in the cloud, with local workflows coming. Internally, Anthropic staff describe it as a higher-level coordinator abstraction with evolving long-lived memory and aggregated status updates via a single controlling Claude (Cat Wu, MikeyK). This is one of the clearer productizations yet of multi-session orchestration instead of just “chat + tools.”
• Google and others are standardizing agent infrastructure around managed harnesses, files, and secrets: Google updated Gemini managed agents with a new Antigravity-based harness plus two notably practical APIs: a Credentials API that keeps secrets out of model context via placeholders and trusted-domain egress proxying, and a Files API for artifact movement and persistent sandboxes. The same release claims up to 30% lower costs and 22% higher cache hits. Meanwhile, Perplexity’s Computer, Base44’s phone-calling Superagent, Google Labs’ family-oriented CC agent, and Meta’s desktop Muse for Mac all point in the same direction: persistent agents with scoped permissions, user-specific context, and asynchronous execution as the default UX rather than an add-on.

Jev and “System One” Classification Models as a New Agent Primitive

• TypeSafe’s Jev dominated discussion as a fast, cheap constrained-output primitive: The clearest pattern in the feed is that builders are treating Jev less as a chatbot competitor and more as a routing / judgment / structured-decision layer inside larger systems. Community reactions emphasize using it for LLM-as-judge, harness routing, subagent creation, and structured outputs, with LangChain noting that Jev is useful precisely because it is not meant for free-form generation. Cloudflare already exposed it via AI Gateway, and open reproductions appeared quickly, including openjev-s with Qwen3.6-35B-A3B + SGLang radix cache and browser demos.
• The technical thesis is “replace prompts with discriminative control flow where possible”: Several posts frame Jev as an “AI if statement” or a generalized classifier for harness logic. Examples include a toy Probably language powered by Jev, a predictive launcher / keystroke oracle, and repeated claims that Jev may be especially strong for reranking, instant routing, and typed extraction (AJ Ratner, dbreunig’s skill, Sydney Runkle’s harness post). The core appeal is familiar to systems engineers: push easy, high-frequency decisions into a small, low-latency discriminative model so expensive frontier models can spend budget on harder reasoning.
• But the compaction discourse showed the limits of classifier-first thinking: A widely shared counterpoint from Theo argues that using Jev for aggressive line-by-line history compaction misunderstands how agent memory, reasoning traces, and cache economics work. His critique is substantive: compaction is not just filtering; dropping hidden reasoning payloads can degrade frontier models; and editing history can be more expensive than leaving it alone because it invalidates cached prefixes. He follows with the stronger framing that the interesting idea is not “better compaction,” but whether future harnesses can abstract away KV caching concerns entirely. That debate is more valuable than the Jev hype itself: it forces clearer separation between classification, memory management, and reasoning preservation in agent runtime design.

OpenAI’s Astra Expansion, Legal Verticalization, and Autonomous Capability Demos

• Astra for Law is OpenAI’s strongest vertical packaging move in this batch: OpenAI launched Astra for Law, with 26 partner-built plugins and 47 community plugins and initial rollout through Trusted Access in ChatGPT and Codex, with API access coming later. Vals says OpenAI’s reported runs show Astra for Law beating generic GPT-6 Astra + web search on its legal benchmark at every price point. The packaging matters more than the benchmark delta: OpenAI is turning frontier capability into domain-specific products with maintained configs, tools, and safety defaults rather than leaving verticals to prompt-engineer from scratch.
• Astra also keeps showing up in unusually broad long-horizon evals and demos: Community reports claim GPT-6 Astra beat Factorio: Space Age, outperformed Fable on RollerCoaster Tycoon 2, and was used for codebreaking-style tasks including WWI/WWII German radio messages. Separately, OpenAI shipped Codex voice from phone via GPT-Live-1, Appshots on Windows, and usage analytics for tasks/subagents/chats. Together these paint a fairly coherent product arc: Astra as the reasoning core, Codex as execution substrate, and increasingly rich interfaces for multimodal capture and async orchestration.

Multi-Agent Research, Evaluation, and AI-for-AI-R&D Measurement

• Research harnesses are getting more explicit, modular, and benchmarked: Google’s DeepMind published Stellar Colosseum, a model-agnostic many-agent harness for mathematics and TCS that separates strategy, decomposition, subproblem solving, and verification; claimed results include a Codeforces 4263 and 71.0% on TCS-Bench. NVIDIA-associated work on Agora uses Git commits as shared memory for 13 workers over 12 days, achieving reproducible progress on model initialization without gradient updates. LangChain shared practical lessons from a 200+ tool paid media agent. The common trend is away from vague “agent swarms” and toward explicit memory structures, decomposition patterns, and reproducibility.
• Anthropic published unusually concrete internal metrics on AI-driven R&D: In a notable transparency move, Anthropic released three measurements for tracking AI development: how much AI R&D is done by AI, how well agents are overseen, and how compute is allocated. Secondary discussion highlights striking numbers: Claude-led share of model R&D tasks rising from 1% to 26% in ~6 months, >90% of model R&D work involving Claude collaboration/leadership, and ~30,000 internal agents active. Even if one treats those figures cautiously, this is one of the few public glimpses into AI-lab internal automation as an empirical object rather than a vibes-based argument.
• Benchmark skepticism is becoming first-class: Epoch launched Benchmark Reviews with 15 audits labeled Verified / Flawed / insufficiently documented, and others noted implications such as artificial ceilings from false negatives on saturated benchmarks (nrehiew). Vals introduced Vibe Code Bench 1-100 to measure iterative modification robustness rather than first-pass success. This is healthy: the field is finally spending public attention not only on scores, but on whether the test itself deserves to exist.

Security, Control, and Misalignment: From Exploit Chains to Reward Hacking