KPAI Weekly: Issue 44

Let’s get started—plenty to cover this week.

Anthropic had the biggest two week run across all NVIDIA5 companies. A near-trillion-dollar raise, an IPO filing, and a new flagship model all landed in the same stretch. Here's the recap:

Anthropic also signed a global alliance with KPMG, putting Claude in front of its 276,000 staff and their clients, and published a new report showing how coding agents are being used in the social sciences.

Who’s moving, who’s stalling, and who’s climbing: Ordered by production this week.

Difficulty: Advanced

A lot is going to change in the field of AI over the next few years, and in truth it is near impossible to predict. However, I do have a theory that I’ve run by some very smart people on what the future of work looks like. This is what’s dubbed an AI-first company, and we’re already seeing it in practice in different capacities across Silicon Valley and even locally here in Cleveland.

Here are the 5 layers, and one governance system, that make up an AI-First company.

Layer 1: The Surface. This is where work happens, when you are talking to your AI, how are you doing it? In the ChatGPT App, in your laptop Terminal, a CLI, or are you using an OpenClaw and communicating solely through text message or Slack. The surface is where your comms happen with your AI.

Layer 2: The Model. This is what’s in play behind the scenes across a few layers. What AI model is powering your company brain, your chats, and your agents. This will likely change over time based on costs, but an example is Opus 4.8 via Claude or GPT 5.5.

Layer 3: Memory/Brain. This is the most important layer in the system, outside of the Governance which we’ll get to. The memory system takes in knowledge from your chats, your work, your SOP’s (Standard Operating Procedures), and your data, and forms a unified “brain” fro your company to operate out of. I wrote a long Linkedin post about Memory layer today. Right now CLI’s like Claude Code and Codex are the best way to get knowledge into your memory system quickly.

Layer 4: Execution. This is how work gets done. Right now in a company that rarely uses AI, humans act as executors and the Memory layer above is simply domain knowledge in people’s heads. Over time, the execution layer becomes more agentic with a mix of human and agent actions. Agents understand how to operate because they pull from the SOP’s and skills in the Memory/Brain layer.

Layer 5: Tools. This is where most work gets done, agentic or not. This is your Gmail, your Hubspot, your Sharepoint. Tools that help you execute. Over time tools will adapt to work better with agents as the executors instead of just humans, and we’re already seeing this with companies like Slack or Linear who adopted MCP protocols to talk directly to agents.

Governance. This isn’t a layer in itself but something that ties it all together. It is the key stack that manages what Agents and operators can touch. It handles permissions across the company so data is protected and agents can’t run wild. This is the layer that prevents your AI bill from skyrocketing and ensures productivity.

That’s it. I think top AI companies will try and expand across models, and I think the memory systems of today (Github [code] Repositories) will sure up into clean products, but at the end of the day I think this architecture stays the same.

Here’s an example of a memory system (OpenBrain) ⬇️

Interview Outline: Salim Ismail breaks down the concept of the "Organizational Singularity," explaining why agentic AI completely breaks Ronald Coase's 1937 theory of the firm. He introduces the ExO 3.0 model—an AI-native destination architecture focused on organizing around intelligence rather than traditional management hierarchies. Ismail details the "Fiduciary Wedge" gap between human liability and machine execution, a 10-week methodology for bypassing corporate immune system responses, and the specific "Rewrite" framework for spinning up an autonomous digital twin at the edge of an organization.

About the Interviewee: Salim Ismail is a prominent technology strategist, entrepreneur, and the co-author of the seminal business books Exponential Organizations and ExO 2.0. He acts as the strategic head of OpenExO, a global transformation community spanning 50,000 practitioners across 150 countries.

Interesting Quote: "All of our organizational structures in the past were organized around hierarchy, and now they need to be AI-native agentic workflows. It needs to be architected around intelligence, not around hierarchy."

Linear — The project management tool built to handle the "Agentic Era."

The Problem: Traditional project management platforms like Jira or Asana are slow, cluttered, and bloated with endless customization. They were built for corporate scrum masters, not fast-moving builders. Even worse, as companies start deploying autonomous AI agents to write and review code, traditional tools break—they aren't built to let humans and AI agents collaborate on tasks seamlessly.

The Solution: Linear is an incredibly fast, design-first product development system built specifically for software teams. Instead of forcing teams into complex, manual tracking, it focuses on hyper-speed keyboard shortcuts, automatic syncing, and clean workflows. Crucially, they’ve recently refactored the platform with a built-in sync engine purpose-built to let human engineers and autonomous AI code agents seamlessly assign, track, and collaborate on tasks together without stepping on each other's toes.

The Backstory: Founded in 2019 by Karri Saarinen and Tuomas Artman (alums of Airbnb, Coinbase, and Uber), Linear has become a cult favorite among elite tech teams. They became legendary for extreme efficiency—hitting over $100M in revenue with under 180 employees. Valued at $1.25B following their Series C, their customer list is a "who’s who" of frontier tech, including OpenAI, Perplexity, and Scale AI.

My Thoughts: Heard a lot of recommendations for this platform as of late. I’m going to start using it in my company in the near future. Seems like a way to handle programming tasks and tickets in a somewhat autonomous and agent friendly manner, while also enabling connection to other outside tools or systems via MCP.