Agentic Companies: Open Source Models Is All You Need

In my recent posts, I’ve explored how private models are becoming more important than massive, closed ones and why the future of agentic AI lies in giving them simple, powerful tools like code and the command line. The release of OpenAI’s GPT-OSS is the final piece of the puzzle, accelerating these trends toward their logical conclusion: the rise of the Agentic Company.

This isn’t just a company that uses AI. It’s a company that is AI. An organization where core operations—from software development to customer support to financial analysis—are not just assisted by agents, but are executed by a cohesive, autonomous system of them.

And the engine for this revolution won’t be a proprietary, cloud-based model. It will be a powerful, open-source model like GPT-OSS, running on private infrastructure.

The Agentic Company: One Organization, One Brain

For an organization to function as a single, intelligent entity, its components must operate from a shared context and a consistent model of the world. This is why I believe we will see the emergence of companies that run almost entirely on a single, fine-tuned base model.

Imagine a specialized version of GPT-OSS, trained on a company’s private codebases, internal documentation, customer interactions, and financial data. This single ‘brain’ powers different agents with specialized roles, but they all share the same core understanding and reasoning capabilities. The engineering agent understands the product roadmap defined by the strategy agent. The marketing agent crafts campaigns based on the features being built by the engineering agent. The system is coherent.

This is impossible with a patchwork of different models or by relying solely on a public API. You need a unified, customizable foundation—a company ‘OS’—and that’s what open models provide.

The Primitives of Action: Code and Browser

But a brain is useless without hands. How does an Agentic Company act? OpenAI’s gpt-oss release provides a profound answer not just in the model itself, but in the reference tools it ships with: a python interpreter and a browser.

This isn’t an oversight; it’s a philosophy. While proprietary agents like Anthropic’s Claude offer a suite of curated tools—a Bash tool, Text editor tool, Web search tool, etc.—gpt-oss provides something far more fundamental.

• Python: This isn’t just a ‘code execution tool.’ It’s a universal solvent for automation. An agent with a Python interpreter can write code, interact with databases, call any API, manipulate files, and perform complex data analysis. It can recreate the functionality of a Bash tool or a Text editor on the fly, tailored to the specific task. This is the engine of product development and internal operations.
• Browser: This provides a window to the world’s real-time information. It’s the engine for research, market analysis, and contextual customer service.

This approach contrasts sharply with the proprietary ‘curated toolset’ model. A curated set is safer and simpler, but it’s also a walled garden. The gpt-oss approach gives the agent the fundamental building blocks of digital work. For an Agentic Company, this is critical. You don’t want an agent that can only use a vendor-supplied text editor; you want an agent that can use python to script vim or VS Code exactly as a human engineer would. This is the difference between giving someone a fish and teaching them how to build a fishing fleet.

The Privacy and Control Mandate

The power of these fundamental tools makes the need for privacy and control even more acute. Granting an AI agent the ability to execute arbitrary Python code is immensely powerful, but it’s a capability you must own completely. Running it on-premise is the only way to ensure it interacts solely with your internal systems, databases, and infrastructure without security risks.

This is why the most innovative applications will be built on private models:

• Healthcare: An agentic hospital system could use a Python tool to query patient databases and a browser to research the latest medical journals, but only if that activity is guaranteed to be secure and compliant.
• Government & Defense: An agency using AI for intelligence analysis cannot risk its browser or code execution tool leaking information. The entire system must run on secure, sovereign hardware.
• Finance & Legal: Firms need agents that can execute complex financial models in Python and research case law, with absolute certainty that this activity remains confidential.

OpenAI itself acknowledged this need, highlighting how these models ‘empower everyone… to run and customize AI on their own infrastructure.’ For the most sensitive and high-stakes use cases, this isn’t just a feature; it’s the entire value proposition.

The On-Premise Revolution

This leads to a radical vision for the future of enterprise: the fully on-premise, air-gapped company.

Think of a business whose entire cognitive workload runs on a private cluster of hardware—maybe even a rack of iPads or custom devices co-located in a secure facility, completely disconnected from the public internet. This isn’t just a theoretical concept. As I explored when discussing the challenges of deploying LLMs on private infrastructure, the tools to make this happen are being built now.

The release of GPT-OSS, which can run on a high-end laptop, proves that state-of-the-art performance is no longer chained to massive data centers. This will fuel a new generation of research and startups that are, by design, private and sovereign. They will be immune to API deprecations, sudden price hikes, and the prying eyes of model providers. They will own their intelligence, end-to-end.

The pieces are all on the table:

1. Powerful, efficient open models (GPT-OSS).
2. Fundamental, unconstrained agentic tools (Python, Browser).
3. Maturing on-premise deployment solutions.

The founders who assemble these pieces won’t just be building another AI-powered SaaS tool. They will be building the first true Agentic Companies, a new corporate structure for the age of AI. The only question is, what will you build?