Snowflake Just Bet $6B on AWS Chips. Nvidia Is Watching.

Remote just told the world it grew revenue 50% per employee, and didn't add a single person to do it.

That number is worth pausing on. $300M ARR, cash-flow positive, same headcount. The story attributes it to AI adoption across internal workflows. No layoffs announced, no drama, just productivity compressing what used to require more people.

This is the version of the AI-labor story that's harder to argue with. Not "AI took my job." More like: "The team that used to need 10 people delivered the same output with 7, and management noticed." Remote is a payroll company. They process employment for tens of thousands of businesses. If this is what their internal stack looks like after AI adoption, it's a preview of what their customers are about to do.

Meanwhile: Snowflake signed a $6B deal with AWS to lock in chips for AI workloads. Robinhood opened its trading platform to AI agents. And a critical vulnerability in a package with 325 million weekly downloads is sitting in your agent stack right now if you haven't patched.

Today's issue runs through all of it.

---

1. Remote hits $300M ARR with no new headcount, AI absorbs the growth Payroll platform Remote surpassed $300M ARR and turned cash-flow positive. Revenue per employee grew 50%, all attributed to AI tooling adoption. No new hires required. TechCrunch

2. Snowflake signs a $6B deal with AWS for AI chip access Five-year agreement, AWS infrastructure, CPU/chip allocation for AI workloads. Nvidia gets a clear message: hyperscalers are routing around single-vendor dependency. TechCrunch

3. Robinhood opens AI agent trading accounts Traders can now create a dedicated account for an AI agent, fund it with a set amount, and let it trade autonomously. Gains and losses are real. Robinhood says this is intentional. The Verge

4. "BadHost", a critical flaw in Starlette, 325M downloads/week Security researchers found a critical vulnerability in Starlette, a foundational Python web framework used in millions of AI agent deployments. If your agent stack touches FastAPI or any Starlette-based service, check your version. Ars Technica

5. Simon Willison: Anthropic and OpenAI have found product-market fit First profitable quarter for Anthropic is apparently close. Companies are reporting surprising LLM bill sizes from internal staff usage, not from products, from their own people using these tools every day. That's what PMF looks like from the inside. Simon Willison

---

[Repo] repomix, Packs an entire repository into a single LLM-ready file. When you need to feed a codebase to Claude or GPT without hitting context limits from file-juggling, repomix outputs one clean file. 8.5k+ stars, actively maintained, works with any model.

[Repo] filetree-skill, A Claude Code plugin that maintains a live FILETREE.md in your repo. Every time your project structure changes, the plugin rewrites the map. Practical fix for the common problem where Claude loses track of where things live across long sessions.

[Repo] opencode, The open-source terminal-native coding agent Dax Raad built and shipped publicly this week. Built on SST, runs in your terminal, supports multiple model backends. The Pragmatic Engineer breakdown covers the architecture decisions and where Dax thinks coding agents hit their limits.

[Repo] microsandbox, Lightweight sandboxed execution environment for running untrusted code from AI agents. Scored 14 unique clicks last issue when it ran, still relevant because the Starlette "BadHost" news above makes isolated execution a live concern, not a nice-to-have.

[Repo] gptme, A self-hosted, terminal-first LLM assistant with tool use, code execution, and file editing built in. No API required for local models. Runs Claude, GPT, or local backends. Good fit for operators who want an agent loop they fully control without a managed service sitting in the middle.

[Skill] SSD fingerprinting via browser JavaScript, Researchers demonstrated that websites can infer visitor identity by timing SSD I/O patterns through standard JS, no permissions required. The technique exploits predictable read/write behavior differences between drives. Worth knowing if you're building anything browser-facing that handles sensitive sessions. Ars Technica

[Skill] Self-improving agent loops with Codex, OpenAI published the architecture behind a tax agent that rewrites its own prompts based on filing outcomes. The pattern: run, evaluate output quality, flag weak steps, regenerate those prompt segments, repeat. OpenAI Blog Same loop applies to any domain where ground truth exists after the fact.

---

[MCP] Vercel Sandbox MCP, Vercel sandboxes now persist filesystem state between sessions by default, no snapshots required. This matters for any agent that writes files mid-task: the state is there when the agent picks back up. GA as of this week. If you deploy previews as part of a pipeline (relevant to tomorrow's kit), this closes a common stateless-session failure mode.

[Tool] Cisco + OpenAI Codex for enterprise defect remediation, Cisco is running Codex to automate defect identification and patching at scale. The non-obvious move here: they use Codex to classify defect severity before routing to human engineers, not just to write the fix. Classification step is the use point. Haiku does the same thing at micro-scale for the kind of pipelines this newsletter covers.

[Tool] Fireworks AI inference, Fireworks just crossed decacorn valuation with a bet on fast, cheap inference at scale. If your current pipeline is bottlenecked on latency or cost at volume, Fireworks is now a serious alternative to direct Anthropic/OpenAI APIs for non-frontier tasks. Worth benchmarking on your classification and extraction steps.

[Tool] Vercel Deployments List, redesigned, Denser layout, environment grouping, branch-level status visibility. Small UX change with real workflow impact: if you're running per-prospect preview deploys at any volume, the new layout makes branch status legible at a glance without clicking into each deployment.

---

Pattern: Use Haiku as a classification gate before routing to Sonnet

Most builders default to Sonnet for every step in an agent pipeline. The cost creep is real and usually invisible until the bill arrives.

The fix is a classification gate: run a cheap Haiku call first to score or label the input, then route only the cases that need it to Sonnet. This week's OpenAI/Cisco example does exactly this, Haiku-equivalent classification before the heavier model handles the actual work.

The command pattern looks like this in any orchestration layer:

const label = await haiku.complete(classifyPrompt(input)); if (label === "needs_render") { return await sonnet.complete(renderPrompt(input)); }

At $0.0001 per Haiku call, you can classify thousands of inputs before the cost registers. OpenAI's self-improving tax agent uses the same structure. OpenAI Blog reference

---

Remote's 50% revenue-per-employee jump is the cleanest data point so far on what AI adoption actually looks like inside a company, not layoffs, not announcements, just compressing the denominator quietly while the numerator keeps growing. The companies that will feel this first aren't the ones building AI. They're the ones whose customers are hiring companies like Remote to manage payroll for the humans they haven't replaced yet.

---

Tomorrow's full kit is the AI Website Agency for Small Businesses. The idea: an AI pipeline reads a prospect's existing site, extracts their brand identity, renders a 10x-better version in their own colors and voice, deploys it live, and sends them a cold pitch with a link to the site they could have, before they've paid a dollar. Pricing is $497 setup plus $99/month per client. The math on who you'd target, why the close rate works at those numbers, and which niche to start with is all in the Friday drop. Get early access at theaigent.xyz/upgrade.

---

The AI-labor story isn't coming, it's already in the quarterly filings.

See you tomorrow.