Mission #14: Agents are everywhere. Their tools are not.
Agents are everywhere now. What's rare is what they can call. When Claude or Cursor hits a real task, like redacting PII from a contract, parsing a quarter's invoices, or searching hours of footage, the reasoning is there. The tool usually isn't. Building that tool takes someone who has shipped for companies and knows where these things break.
Here's what Cloud does with that: deploy your pipeline, and it's a live MCP tool on an always-on endpoint. No wrapper, no server, and more than a hundred nodes to build with. Any MCP-compatible agent can discover it and call it, at 2am, under load, while you're building the next one. On your laptop, your tool keeps your hours. On Cloud, it works for the whole ecosystem.
Everyone is racing to build a smarter agent. The better position is quieter: be the toolmaker every agent depends on.
"A guardrail that locks out the attacker locks out the incident responder too." That's one of the sharper observations in Mithilesh's new security post, drawn from the Hugging Face team's own account. When they tried to analyze the attack data, commercial models refused the work, so they ran forensics on a self-hosted open-weight model instead.
The post walks through the whole incident, IBM's 2026 breach numbers, and the three defenses that actually matter in production: containing what a foothold inherits, keeping credentials out of the pipeline artifact, and seeing inside a run while it's running. Where a runtime stops and your judgment starts is the part worth reading twice.
The week's AI news that changes how you build.
Nvidia just made "tokens per second" its newest product line.
On Monday Nvidia announced full production of Groq 3 LPX, a dedicated inference accelerator for its Vera Rubin platform, built on LPU technology licensed from Groq. The headline number: 3,400 output tokens per second on Gemma 4 31B, with 100K context. Nvidia claims that is 4x faster than the nearest alternative for latency-sensitive agent workloads. Nebius signed on as the first cloud to deploy it. Every multi-step agent you run has a hidden budget of "how long will a user wait", and the ceiling on that budget just moved.
Vision just got cheap enough to leave on in every pipeline.
DeepSeek shipped V4-Flash-Vision-Exp on Friday: the same text and agent capabilities as V4-Flash, now with image understanding. Images are billed as ordinary tokens, capped at 384 tokens per image at Flash pricing. DeepSeek's own benchmarks put its multimodal agent performance nearly level with Opus 4.8, with the usual caveat that it is an experimental model and those are vendor numbers. A free Files API launched alongside it, so you upload an image once and reference it by file ID across requests. If you cut a screenshot-reading or document-vision step from your pipeline because the quote was ugly, this is the week to re-price it.
Open models just became the majority of production traffic.
Vercel CEO Guillermo Rauch published AI Gateway data this week showing open-weight models jumped from 28 percent to 62 percent of token share in two months. OpenAI and Anthropic are still growing in absolute volume; what flipped is where new workloads go, with cheap open generalists taking the bulk and closed frontier models keeping the hard cases. That is production traffic across real apps, not a benchmark. If every node in your pipeline still defaults to one closed model, the market just showed you what your bill could look like instead.
Member discussion