New MLPerf® Storage v3.0 results are in! Nebius Object Storage fed 768 simulated B200 accelerators on RetinaNet, the highest accelerator count of any submission in this round. Get the full breakdown: #MLPerf# #AI# #ObjectStorage#

What does it take to run object storage without a single traditional server? Garage is a lightweight, S3-compatible object store. Using Acurast Cargo, it runs entirely on decentralized compute. Same S3 API. No servers to manage. Try it out yourself:

Loop is now so good at detecting patterns that we're using it to optimize Brainstore. We pipe all of our query plans into Braintrust and found a bunch of low-level optimizations, eg a spot in our regex code where we were over-fetching duplicate ranges from object storage.

Today we're launching Agent Bucket: native storage for AI agents at scale. Standard S3 API, so you spin up independent space for billions of agents with tools you already know. Object storage was not built for agents. One bucket each hits limits fast. One shared bucket pushes isolation into your app code. So we rebuilt the foundation. → Independent space at scale. Each agent gets its own Space, with dedicated credentials, quota, and rate limits. Isolation becomes a system capability, not app code. → File management out of the box. Dedup, POSIX semantics, multimodal search, sharing links, recycle bin, content moderation. All at the storage layer. → Fast and low cost. GooseFS acceleration keeps hot data close to compute, with hot and cold tiering to balance speed against cost. → Proven at scale. Already running inside Tencent's own agents WorkBuddy, QClaw, and across external use cases such as Asus AI-PC drives, connected cars.

We’re introducing Backup and Restore for OCI Cache. A new capability that allows customers to: ✅ Create on-demand snapshot backups of cache data ✅ Retain backups in OCI Cache for up to 35 days ✅ Export backups to customer-owned OCI Object Storage ✅ Restore backups into a new cache cluster from either OCI Cache or Object Storage

You can spend months building out your Hermes profiles. But if something unexpected happens to your environment, are you actually protected? A bad update, corrupted state, drive failure, or a broken setup can turn all that accumulated work into a recovery problem fast. Memory. Sessions. Skills. Automations. Configs. All the little things that make your setup yours. Backing it all up takes one command: `hermes backup` That creates a full Hermes backup you can restore later with `hermes import`. But I wouldn’t stop at running it once. Automate it. Hermes supports recurring no-agent cron jobs, which means you can have Hermes run a backup script for you on a schedule without firing up an LLM or using model tokens. Hermes can literally back up Hermes. Then get at least one copy off the machine running it. That could be: • Google Drive or OneDrive • Another PC, NAS, or external storage • Cloudflare R2 or similar object storage • An encrypted cloud destination through something like rclone And yes, GitHub can have a place here too, just not the way you might first think. I’d use a private repo for versioning the rebuildable parts of a profile like your SOUL, skills, scripts, and sanitized configs. I would not dump raw full Hermes backup ZIPs into GitHub. Full backups can contain things like `.env`, `auth.json`, sessions, memory, and other private state. Treat those archives like sensitive data. Hermes also supports backup retention with: `hermes backup --keep N` So you can automate the process without letting old backups pile up forever. The exact place you store them matters less than the habit: Make the backup automatic. Keep at least one copy somewhere other than the machine running Hermes. You’ve probably put way too much work into your agents to rebuild all of it from scratch.

Practices for embedding AI agents into enterprise systems [Observability / Tracing + Provenance] 💡 An agent that cannot explain "why it gave that answer" has no place in production. Structured tracing and provenance make probabilistic behavior auditable. 🔥 Problems solved - Cannot reproduce or analyze why a specific answer was generated - No visibility into costs per department, project, or agent - Silent quality degradation from model or prompt changes goes undetected - Regulated decisions lack explainability for post-hoc accountability 🏗️ Proposed pattern Record every reasoning step, tool invocation, token count, cost, latency, and eval score as OpenTelemetry-compliant structured traces. Store metadata in log infrastructure and full prompts/raw outputs in object storage, linked by trace ID. Sample normal requests (1-10% as a starting point) and record all errors and low-scoring responses in full (tail-based sampling). For regulated decisions, extend traces with provenance -- tracing back to source documents, reasoning paths, model versions, and human approvers for full accountability. ✅ Selection criteria - Use when: Every agent in production, no exceptions - Skip when: No exceptions -- this is a mandatory pattern for production ⚠️ Pitfalls - Storing full prompt text in log infrastructure is cost-prohibitive at scale - Failing to mask PII turns trace logs themselves into a security liability - Design decisions on provenance granularity require human review, not autonomous judgment 🛠️ Implementation Approach 1. Instrument all agents with OpenTelemetry GenAI semantic conventions, recording reasoning, tool invocations, and retrieval steps under a consistent trace ID 2. Send metadata (model name, token count, latency, cost, eval scores) to an LLM observability platform (Langfuse / LangSmith / Arize) and monitor per-department cost and quality trends on dashboards 3. Store full prompts, context, and raw outputs in object storage (S3, etc.) linked to log infrastructure metadata by trace ID 4. Implement tail-based sampling: sample 1-10% of normal requests while recording all errors, low-scoring, and high-cost requests in full 5. For regulated use cases, maintain decision logs as append-only immutable audit records with provenance linking back to source documents, model versions, prompt versions, and human approvers #AIAgents# #EnterpriseArchitecture#

More likely than not, you’ve been using AI tools such as Claude Code for day-to-day work. You may have noticed that when AI needs to look at a data file, it makes use of Python, writes a small script, runs it, and then reads the output. It’s an approach that works most of the time, but it’s slow and not always accurate. The duckdb-skills plugin gives Claude Code a growing number of skills that use the DuckDB CLI to read data files, run queries, convert formats, explore object storage, work with spatial data, search the documentation and recall earlier sessions. Here’s how you can get started with the new plugin:

$CRWV DEPLOYS MULTI-RACK $NVDA VERA RUBIN NVL72 CLUSTER CoreWeave says it has connected hundreds of Nvidia Rubin GPUs into a single scale-out cluster on CoreWeave Cloud, targeting large agentic AI workloads. Each Vera Rubin NVL72 rack combines 72 Rubin GPUs with 36 Vera CPUs, while the multi-rack setup uses Nvidia Spectrum-X networking to link racks together. Each Rubin GPU gets 1.6 Tb/s of network connectivity, with the fabric designed to scale to roughly 128,000 GPUs per rail. CoreWeave also rolled out cross-region write acceleration and a new lower-cost Archive tier for AI Object Storage. Its LOTA storage accelerator claims up to 7 GB/s throughput per GPU and an 8x reduction in read latency versus traditional storage clusters. CoreWeave says it was the first AI cloud provider to validate and deploy a single-rack Vera Rubin NVL72 before expanding to multi-rack deployments.