에이전트 실행 인프라, 연구형 모델, 검증 가능한 시스템, 모델 라우팅의 확장

Box

@Box

2일 전

@OpenAIDevs Agents API gives developers a hosted sandbox where agents run tools, coordinate work, and handle complex tasks. Box Mount brings enterprise content directly into that sandbox as files the agent can read, reason across, and produce new work from, using normal shell commands and file paths while Box Mount handles two-way sync automatically. In our demo, a lead agent mounts a deal room from Box, reads five source files, launches three specialist agents in parallel, and writes four reports back through Box Mount with an initial verdict. When a customer revises the MSA in Box, Box Mount syncs the new version into the same sandbox and the agents reassess automatically, refreshing all four reports while preserving the revised MSA as version two in Box. The result is a shared workspace where people and agents work on the same governed Box content without building custom file-transfer logic, with permissions, versions, and audit history preserved throughout. Box Mount is in private preview. Watch

Agents API is a bigger deal than it seems. OpenAI's bet on making the Codex harness open-source could pay off big time here. This harness-as-a-service idea hasn't hit mainstream yet, but some of the most serious builders are already adopting it to unlock new types of services. Codex being open-source means that many other models (open or closed) are already extremely knowledgeable about the Codex harness. As more services adopt these new agentic architectures, you'll want the compounding effect that comes with that transparency and openness.

OpenAI Developers

@OpenAIDevs

2일 전

Get more out of GPT-6 Astra by revisiting your skills, AGENTS.md, and task prompts. Make skill triggers specific, load guidance when it's relevant, and define what done looks like.

ChatGPT-6 Astra, Fable 5.1 Frontier coding just got 39% cheaper. Devin CLI now has Fusion: - Fable 5.1 / Astra handle planning and review - SWE-2 handles exploration, implementation, and tests, Same quality, far fewer tokens. - Up to 39% cheaper across coding benchmarks. - On some tasks, savings hit 40-46%. I’ve been using it for days and the token savings are already noticeable.

OpenAI Developers

@OpenAIDevs

2일 전

Bring stronger biological reasoning to your research with GPT-Rosalind in the API and Codex. Connect findings across papers and experimental results, weigh the evidence for a biological target, and work through an analysis to plan what to test next.

check out how AI helps scientists with quantum computing research @bea_yankelevich at MIT's EQuS group used GPT-5.6 Sol with Codex to run routine measurements on quantum chips, so she could accelerate her experiment design

Agents let us build systems for different workloads and requirements. But… can we trust what they build? We release SkySynth: an engine for synthesizing high-performance, just-in-time (JIT) systems we can trust, by co-evolving formal proofs and tests alongside the code. Results: KV stores up to 2.3× faster than Redis and FASTER + formally verified stores with 2.9× Claude Code's pass rate Model routers up to 48% cheaper than a general router Specialized inference engine with 2.2× the throughput of vLLM/SGLang

I've been trying Codex to analyze a dataset, and honestly, I've gotten nowhere. Generating a nice-looking report about the data isn't enough. For this to be useful, Codex needs to inspect tables, clean existing data, choose the correct statistical methods, run the analysis, and support its conclusions with evidence. There are too many tasks to coordinate, and it's taking me a long time to write a skill with the complete workflow so I can keep Codex on task. The Apodex team reached out to me to show me their Apodex 1.1 model. It includes an open-source framework and an open-weight version of the model. I uploaded a CSV file and asked a question that I couldn't answer by simply summarizing the table. Apodex created an analysis plan, wrote the code, and executed it inside a single long-running task: 1. It inspected the tables 2. It determined the steps to answer the question 3. It wrote Python code to clean and transform the dataset 4. It executed the scripts in its environment 5. It generated a final report with the answer This is what Apodex calls Environment Scaling. I could intervene while a task was running, and the system preserved all intermediate results that remained valid and replanned only the affected parts instead of restarting the task. Apodex 1.1 takes reasoning out of the report and puts it into executing real tasks. Thanks to the team for collaborating with me on this post. GitHub links below:

Tailscale offers instant access to hundreds of models for any user in a secure tailnet. Their customer-facing model router, Aperture, is built on AI Gateway: • Zero data retention • Zero markup, free BYOK • Cost and usage data on every response

Guillermo Rauch

@rauchg

2일 전

Tailscale’s model router is powered by Vercel AI Gateway as its underlying infrastructure. AI Gateways are the new CDNs. You could go “direct to origin”, but it’s brittle. You could DIY, but it’s painful and costly. Proud to serve Tailscale, the networking technology company

Google Gemma

@googlegemma

2일 전

Running AI in space requires strict power and bandwidth management. Satlyt is deploying Gemma directly on satellites to analyze telemetry and diagnose faults locally. Running Gemma 3 1B onboard cuts diagnostic data payloads by over 64%, turning raw system logs into compact summaries before they are sent back to Earth. Next launch target: Gemma 4 E2B on NVIDIA Jetson Orin Nano to provide further reasoning and vision capabilities at the edge… of space.

NVIDIA AI

@NVIDIAAI

2일 전

From Video to Voice: Build Faster with TensorRT Model Connect

We analyzed how similar model responses were across 30,086 Arena battles. Models shared 43% of their ideas on average. We might expect that models from the same lab, or country, would show greater conceptual overlap. But the results don’t consistently support that. Claude Fable 5 illustrates this pattern: its closest conceptual match was neither Opus nor Sonnet. Which model came closest, along with the broader findings, may surprise you. Check out the full article from @DawidGalarowicz and @petergostev below.

Frontier models still struggle to truly understand biomedical images... Researchers at Stanford are hosting a much-needed challenge to benchmark and improve biomedical visual understanding. Definitely check it out!