CoreWeave tops MLPerf as AWS doubles down on agents

TLDR: CoreWeave posts best-in-class MLPerf inference results while AWS quietly builds the agent stack, from Bedrock AgentCore to frontier agents.

CoreWeave hits top MLPerf inference scores across AI cloud rivals

CoreWeave reports leading inference performance in the MLCommons MLPerf v6.0 benchmark and is currently the only AI cloud with a Platinum ranking in both SemiAnalysis ClusterMAX 1.0 and 2.0 as of 2026-04-02. The company is framing this as validation that its GPU-centric cloud and software stack are purpose-built for production AI workloads rather than generic compute.

For anyone running high-throughput inference or planning multi-tenant agent backends, these results suggest CoreWeave remains one of the most performance-focused alternatives to hyperscalers. Benchmarks are synthetic and do not capture pricing, networking quirks, or long-tail failures, so treat this as signal on raw capability, not total cost of ownership. If you are pushing for aggressive latency or token-per-dollar targets, this is a data point worth bringing to your infra team.

Read more →

---

AWS details Bedrock AgentCore stack for enterprise agentic commerce

Amazon Web Services is publicly positioning Amazon Bedrock AgentCore as its opinionated platform for agentic commerce, built around a consistent set of building blocks for agents as of 2026-04-02. AgentCore wraps enterprise identity, inbound and outbound authentication, runtime orchestration, tool and data access controls, memory, evaluation, and end-to-end observability so teams can safely deploy agents that take multi-step actions, not just generate text.

If you are building transactional agents in regulated environments, this is effectively AWS saying: do not hand-roll identity flows and observability; use our primitives instead. The value is clearer for enterprises already standardized on AWS IAM and CloudTrail, and less compelling if you are multi-cloud or deeply invested in open orchestration frameworks. Expect tighter coupling to Amazon Bedrock services, which simplifies integration but can increase lock-in.

Read more →

---

Model Context Protocol bridges AI agents and hotel ops systems

Hospitality Net outlines how Model Context Protocol (MCP) is being used in hospitality as a translation and orchestration layer that sits alongside hotel systems like property management systems (PMS), central reservation systems (CRS), and customer relationship management (CRM) as of 2026-04-02. MCP receives structured requests from AI agents, routes them to the appropriate operational platform, then translates responses into AI-friendly formats without becoming a system of record.

For teams building vertical agents in legacy-heavy domains, this is a clean reference pattern: keep core systems untouched, put MCP-style middleware in the middle, and let agents speak a domain-neutral protocol. The tradeoff is another layer to operate and secure, but it limits blast radius and lets you iterate on agent behavior independently of fragile backends. If you are pitching agent projects to IT groups who fear data silos, this architecture gives you shared language.

Read more →

---

Quick Hits

• Multimodal Embeddings and RAG: A Practical Guide This Weaviate tutorial walks through how to use multimodal embeddings for retrieval-augmented generation (RAG) across text, images, audio, and video, with concrete Gemini plus Weaviate examples.

• Accelerating software delivery with agentic QA automation using Amazon Nova Act AWS shows a reference "QA Studio" that uses Amazon Nova Act to define UI tests in natural language, then runs them on a serverless architecture that adapts to UI changes.

• AWS launches frontier agents for security testing and cloud operations AWS Security Agent for on-demand penetration testing and AWS DevOps Agent for operations are now generally available, targeting long-running, autonomous security and ops workflows with minimal human oversight.

• Building an AI powered system for compliance evidence collection An AWS architecture guide for automating compliance evidence collection with AI agents and managed services, aimed at teams drowning in SOC 2 or ISO audit prep.

• langchain-core 1.2.24 The core package adds placeholder filename support for OpenAI file inputs, updates Pygments to address CVE-2026-4539, and expands well-known OpenAI tools.

• langgraph 1.1.4 The new release improves recursion limit handling, adds LangSmith integration metadata, and bumps security-sensitive dependencies. Also covered by: github/langchain-ai/langgraph.

• langchain 1.2.14 LangChain ships a 15 percent init speed improvement, fixes recursion limits in create_agent, and standardizes on Pygments 2.20.0 plus newer cryptography libraries.

• TRL v1.0: Post-Training Library Built to Move with the Field Hugging Face TRL graduates from research code to a stable post-training library with clearer guarantees, making RLHF and preference optimization more production-friendly.

• datasette-enrichments-llm 0.2a0 Simon Willison updates his Datasette plugin so enrichments now route through datasette-llm, which centralizes model configuration and lets you scope models by enrichment purpose.

• Holo3: Breaking the Computer Use Frontier Hcompany's Holo3-122B-A10B model hits 78.85 percent on OSWorld-Verified for desktop computer use tasks, setting a new benchmark for autonomous computer use agents.

• BetterDB: open-source Redis or Valkey migration A new tool born from Redis ecosystem fragmentation that migrates data across Redis, Valkey, and cloud providers so you can escape managed lock-in and protocol breaks.

• Claude Code Unpacked: A visual guide A community-built visual walkthrough of the leaked Claude Code source, useful if you are studying how Anthropic wires tools, regex-based guards, and "undercover" modes in real-world agent code.