Private platforms vs. agent-native products

Giving employees a safe chat window was the 2024 project. Wiring agents into your systems is the 2026 one. Most mid-market companies are still finishing the first.

Third post in the series. The figures are the companies’ own public claims, unverified by me, and wildcard* had no part in any of them.

Start with the platform generation. JPMorgan Chase built LLM Suite, a private platform on OpenAI models inside the bank’s own infrastructure with strict data isolation. By the bank’s earnings transcripts it reached more than 60,000 employees, used for research and email drafting. Huawei and Imperial College, per an SG Analytics report on DeepSeek, went further on control: self-hosted open-weight models on Imperial’s dAIsy platform, so no data leaves the premises and the model is fully owned.

Both solved the same problem: people wanted to use AI, security said no, so the company built a walled garden. The output is text that a human copies somewhere else.

Now the agent-native generation. Klaviyo, in Anthropic’s customer stories, runs Claude Code with repository context in an AGENTS.md file and Model Context Protocol connections to live customer data. The agent creates campaigns grounded in real data and maintains SDKs. Figma, same source, turns natural-language prompts and design concepts into interactive components, cutting the handoff between design and code. Chinasoft and Kingsoft, in Moonshot AI’s enterprise case studies, deploy agent swarms that run multi-step business workflows across separate enterprise systems.

The difference is where the output lands. A platform produces a draft for a human. An agent-native setup produces a change in a system: a commit, a campaign, a completed workflow.

For a DACH company with 30 to 300 engineers, here is what I would copy and what I would skip.

Copy the AGENTS.md. One file in each repository that tells an agent the conventions, the test commands and the boundaries. It costs an afternoon and raises the quality of every agent interaction. Klaviyo’s setup is not exotic. It is written-down context.

Copy MCP, or whatever protocol your tools support, for read access first. An agent that can read your CRM, your ticket system and your database answers questions nobody dared to ask before. Write access comes after you have evals.

Copy the data isolation principle, not the implementation. You do not need a bank’s private platform. You need a contract with your provider that says where data goes and a policy that says what may be pasted.

Skip the 60,000-seat platform. You have 100 people. A managed enterprise plan and a one-page policy give you the same control.

Skip agent swarms until one agent does one workflow reliably, measured. Chinasoft and Kingsoft are system integrators with large teams behind the swarm. Your first agent should be boring.

The order matters: context file, read access, evals, one write path, then more. Most companies want to start at the swarm.

Where does this break? Tell me about a company at your size that got value from the platform alone.