NVIDIA Assembles the Agent Coalition

Edition 167 | March 19, 2026

SF texting agent Poke has an aggressive price haggling process where you try to convince the agent how much you should be paying for it. A lot of fun to see how hard you can push the agent. I got mine down to $15/mo from $200/mo.

Just to be clear - this is not sponsored. You can text it without paying!

Welcome back to Building AI Agents, your biweekly guide to everything new in the field of agentic AI!

In today’s issue…

• NVIDIA assembles open-source coalition to build agent-native models

• Manus launches desktop agent with full local file access

• Appeals court lifts block on Perplexity's Amazon shopping agent

• Perplexity ships Comet, an agentic browser for iOS

• OpenAI reframes prompt injection as social engineering

…and more

🔥 INCASE YOU MISSED IT

Readers’ favorite items from the past week

1. Perplexity launches Computer for personal and enterprise use

2. Why Claude Code is the inflection point for AI agents

3. How LangChain built a sales agent that 3x'd its pipeline

4. Karpathy lets an AI agent run his ML research overnight

5. LangChain launches Deep Agents for multi-step workflows

📌 THE THURSDAY BRIEFING

NVIDIA assembles an open-source Avenger’s style coalition for agent models

📡 Signal: NVIDIA launched the Nemotron Coalition Monday at GTC, a first-of-its-kind collaboration to co-build open frontier models built specifically for agent workloads. Mistral AI, Perplexity, LangChain, Cursor, Black Forest Labs, Reflection AI, Sarvam, and Thinking Machines Lab (Mira Murati's new company) are all founding members. The first model, co-developed with Mistral, will be open-sourced and underpin the Nemotron 4 family. LangChain is building the agent harness and eval layer, Cursor is contributing real-world coding benchmarks, and NVIDIA shipped OpenShell for safe agent runtime and NemoClaw for agent security guardrails alongside the announcement.

🤖 For builders: This is NVIDIA's play to own the open-source foundation layer for agents. Right now, most production agents run on closed models from OpenAI, Anthropic, or Google. Nemotron 4 is being built with agent tool use, long-horizon reasoning, and orchestration baked into the training process, not add on after. If you're building agents on open models, this is a big launch to watch.

Meta’s Manus launches their desktop agent

📡 Signal: Meta's Manus launched My Computer, a desktop app for Mac and Windows that gives an AI agent full access to your local files, apps, and workflows. It handles tasks like organizing files, managing emails, and navigating apps on your behalf. The launch is a direct response to OpenClaw's momentum.

🤖 For builders: This is agents moving from cloud dashboards to the machines people actually work on. The shift from "agent in a browser tab" to "agent with access to your filesystem" is a meaningful jump in capability. The question is no longer whether people will let agents operate their computers. It's which agent gets there first.

🤝 WITH FISHER INVESTMENTS

Your Retirement Savings Need to Outlast You

Most retirement plans underestimate two things: how long your savings need to last, and how quietly inflation erodes them along the way.

The 15-Minutes Retirement Plan helps you close both gaps with practical guidance on longevity risk, purchasing power, and building a financial plan that doesn't run out before you do.

If you have $1,000,000 or more saved, download your free guide to start.

Download Your Free Guide.

 🤖 AGENT OF THE WEEK

👋 Welcome back to Agent of the Week!

Every company has sales dashboards. Revenue charts, weekly summaries, month-end reports sitting in some BI tool or messy Google Sheets. The data is there. What's missing is someone to actually analyze it.

Big companies have teams of analysts for that. Most businesses don't. You're either that person yourself, or you're spending an hour clicking through dashboards trying to answer a question that should take 30 seconds.

Claude Cowork changes that. Point it at your sales data (really any data) and have a conversation with it like you would with an analyst sitting across the table. No SQL. No formulas. Just plain English questions, real answers, and visuals that actually answer the question you asked.

I tested this with a client's restaurant POS data. Years of transaction data they weren't using to make informed decisions. Questions like: did that promo actually drive enough sales to justify the discount? How am I doing vs last year for the same month? Are my Fridays picking up and why? These aren't complex questions. But most SaaS tools make them surprisingly hard to answer.

Within minutes I had week-over-week revenue trends, hourly heatmaps showing when the dinner rush peaks and dies, margin breakdowns between food and drink, profitability comparison across dine-in/takeout/delivery, and poor performing menu items that should have been killed a while ago.

This completely changed how the owner was thinking about her operation. Here’s how I set it up for her.

🧠 How It Works

1. Connect Your Data → Give It Context: Load your data from a Google Sheet, CSV, or platform export. Tell the agent what it's looking at. That context is what turns it from a chatbot into an analyst

2. Ask Questions → Keep Digging: Start simple. "What's my most profitable day?" The answer leads somewhere. "Why is Saturday margin lower than Friday?" One question sparks the next. It's a conversation with a coworker, not a one-shot prompt

3. Generate a Visual Report → Charts, Summaries, Recommendations: When you've found what matters, have it build everything into an HTML report. Revenue trends, hourly heatmaps, margin breakdowns, all visualized. A real report you can open and walk someone through

4. Reuse It → Refresh Anytime: Next month, hand Cowork the new data and say "run it the same way." Same structure, fresh numbers. You just built a repeatable analytics workflow with zero code

There's no reason not to have an agent looking at your sales data weekly. It's virtually free if you already have a Claude or OpenAI subscription, and a lot cheaper than burning your team's time pulling reports you could answer yourself in 5 minutes. Just talk to your data, get answers, and show up to your team with sharper questions.

That's the game changer, self-serve business intelligence without being technical or setting up another dashboard.

This works for any business with data. Restaurant, agency, SaaS, ecommerce, etc. If it's stored somewhere, you have a personal sales analyst working for you 24/7.

Having the data was never the problem. It's been having the time to actually dig in. That excuse is gone now.

Till next week,

✌️ AP