Plus: vibecoding takes 100 hours to ship, Ramp runs AI agents at scale, and vLLM gets 1.69x faster inference.
Anthropic is reportedly in talks with Blackstone to form a Palantir-style enterprise consulting JV, Google closed its $32B Wiz acquisition (the largest startup deal ever), and Anduril locked in a $20B Army contract. On the builder side, MCP is not dead — it's more relevant than ever for enterprise agentic tooling. We also learned that vibecoded prototypes take 100 hours to ship (not 30 minutes), and that prompt injection against AI agents is now an infrastructure problem, not a model one.
Google acquired cybersecurity firm Wiz for $32 billion, marking its largest-ever acquisition and the biggest buy of a venture-backed startup. Wiz's position at the nexus of AI, cloud, and security spending provided significant tailwinds for the deal, according to its largest shareholder, Index Ventures.
Read full story→The US Army awarded Anduril a 10-year contract potentially worth $20 billion, consolidating over 120 prior procurement actions. This move highlights the growing reliance on software speed and efficiency in modern defense, with implications for AI integration in military systems.
Anthropic is in talks with Blackstone and Hellman & Friedman to form a joint venture that would sell AI consulting services to their portfolio companies. The Palantir-style model would combine Claude with hands-on consulting to accelerate enterprise AI adoption, giving Anthropic access to Blackstone's $1T+ asset portfolio without building an enterprise sales org from scratch. Talks continue despite Anthropic's ongoing dispute with the Pentagon, which designated it a supply chain risk over its restrictions on military AI use.
Vibecoding an AI app in 30 minutes is hype. Turning that prototype into a deployable product takes 100 hours. The bulk of that time goes to UI/UX, infra, and edge cases. AI accelerates the start, but human engineering nails the final 90% of effort.
P-EAGLE, a new parallel speculative decoding method, accelerates LLM inference speeds up to 1.69x over EAGLE-3 in vLLM on NVIDIA B200 GPUs. It generates all draft tokens in a single forward pass, ditching sequential processing to cut overhead. vLLM users can enable this by setting `parallel_drafting: true` and using pre-trained P-EAGLE heads available on HuggingFace.
A technical breakdown of how Vision Language Models (VLMs) are built, revealing why adapting text models is preferred over expensive from-scratch training. The architecture leverages a frozen Image Backbone, a Q-Former Adapter Layer that uses cross-attention to link vision and text embeddings, and a Language Layer. This compute-efficient method forms the basis of most modern VLMs.
One builder's experience shows that designing products for AI agents demands an agent-first approach, prioritizing programmatic interaction over human UI/UX. Key principles include clear `Skills.md` documentation, API-driven account management, agent-aware rate limiting, and machine-readable errors. This approach allows agents to operate at scale, handle high-volume API calls, and offers a tight feedback loop for developers to iterate faster.
Ramp, valued at $32 billion, is automating operations with internal AI agents handling tasks from customer research to data analysis. Their 'product shaping' Claude Code skill acts as a product thinking partner. The company aims for 80% of its codebase to be AI-generated using its L0-L3 framework, enabling all employees to build with AI.
NVIDIA introduced an agentic retrieval pipeline, achieving the #1 spot on ViDoRe v3 and #2 on BRIGHT leaderboards. This pipeline creates an iterative loop between an LLM and a retriever, using a ReACT architecture for planning and retrieval. Engineering optimizations, like an in-process singleton retriever, cut latency and complexity, allowing dynamic adaptation to diverse data and complex reasoning tasks.
Ankur Sethi built a new programming language, Cutlet, with Claude Code in four weeks, letting the LLM generate all code under verification guardrails. The project demonstrates LLMs can compress six months of development into four weeks. It also outlines four critical agentic engineering skills: problem selection, clear intent communication, environment setup, and loop optimization, suggesting software engineering transforms rather than disappears.
Prompt injection attacks are evolving, enabling agents to access private repos, send emails, or execute code. This shifts the problem from model-level to infrastructure, requiring defenses like scoped permissions and treating tool descriptions as code. Agent security is converging with traditional application security, predicting future agent permissions will resemble cloud IAM.
The buzz around Model Context Protocol (MCP) has moved to CLIs, but MCP remains vital for enterprise agentic tooling. Centralized MCP over HTTP offers richer tooling, simplified auth, and observability that custom CLIs can't match. For organizations building agentic systems, this structured approach is essential for security and management.
An interview with Vivek Bharathi argues AI is commoditizing software development (manual coding), while software engineering, focused on designing automated systems, evolves. The piece notes traditional open-source models diminish as AI generates code, shifting competitive advantages to non-technical areas like relationships and distribution. The discussion suggests professionals will need to prioritize raw technical and cognitive skills and embrace AI tools to remain relevant.
Captain, a YC W26 startup, launched an automated Retrieval-Augmented Generation (RAG) service. It handles data indexing, including OCR and VLM for file types, and provides a vector database. The platform claims over 95% accuracy with minimal maintenance, a jump from 78%.
Context Gateway is a local intermediary that compresses AI agent conversation history before sending it to the LLM. It automatically summarizes past interactions in the background, preventing context window overloads and reducing token costs. The tool installs via `curl` and configures through a TUI wizard, supporting agents like Claude Code and Cursor, and allowing custom summarization models and trigger thresholds.
Cloudflare's new `/crawl` endpoint provides developers with an API to scrape entire websites with a single API call. It renders pages in headless browsers, outputs HTML, Markdown, or JSON, and respects `robots.txt` and AI Crawl Control. This streamlines data ingestion for RAG pipelines and model training, reducing the need for custom crawling infrastructure.