Ammonix exits stealth with System One + System Two specialist agents

Ammonix emerged from stealth with an AI architecture that pairs a fast specialized System One decision layer with a System Two language model for reasoning, aimed at high-stakes specialist agents that can run on local hardware.

SOFTWARE desk — high-stakes domains want agents that are cheap, local, and auditable, not just another giant general model with more parameters.

What the company says the agents are for. The architecture is meant to build specialist agents that learn from a company’s operational data and workflows. Operational data is the record of how the work actually gets done. A workflow is the sequence of steps. The release says that gives fast responses, recommendations a person can audit, control over sensitive data, and predictable computing. Audit, here, means a person can check the recommendation against examples the agent shows. Inside a larger agent system, Ammonix says its agents can sit as a control layer on critical functions, and that this makes those systems safer, faster, and more reliable. Safer and more reliable are the company’s words. This desk did not measure either.

How Ammonix says an agent learns. Unlike a general-purpose model, the release says, these agents follow a few rules. They learn from the organization’s own data. They reach expert-level performance from a few hundred examples. They get better over time by updating their memory. A few hundred examples is the company’s line for how much labeled work it says is enough. Expert-level is the company’s word. This desk did not count the examples or score an expert. Every recommendation, the release says, is fast and auditable, with supporting examples available to check. When an agent does not know, it asks a person instead of guessing. That ask-instead-of-guess rule is the design the company names for high-stakes work. By capturing a skill once and reusing it, the release says, the agents can run on local hardware. Local hardware means the computers a company already runs, not a rented cloud chip for every question. The release says that cuts compute and cost and improves privacy, security, and control. Dramatically is the company’s adverb. This desk did not open a bill or a server.

Four demos, and whose comparison they are. Ammonix names four applications: reading an ECG, collecting healthcare claims, support in a waste-to-energy control room, and applied computer vision. An ECG is an electrocardiogram, the tracing of the heart’s electrical beat. A claim is a bill a clinic sends to an insurer. A control room, here, is the place operators watch a plant. Computer vision means software that reads pictures. In those four, the release says, the agents matched or outperformed OpenAI’s GPT-6, at lower cost and at significantly faster speeds. GPT-6 is the OpenAI model named on this page. Matched or outperformed, lower cost, and significantly faster are Ammonix’s sentences. The page does not print a score, a dollar figure, or a time. Do not invent them. This desk did not rerun the four jobs.

Who is in charge, and the prior sale. The company is led by co-founder and chief executive Peter Ruppersberg. The release calls him a scientist and a serial entrepreneur. It says his medical AI company, Cortex, was acquired by Boston Scientific in January 2025 for up to $300 million. Up to $300 million is a ceiling on that deal, as the release states it. It is not a check this desk saw, and it is not a price for Ammonix. Do not invent a valuation for Ammonix. The page does not print one.

The quote, as the company’s, not as a result. Ruppersberg said enterprise AI cannot simply be another expensive model with more parameters. A parameter is one learned number inside a model. More of them is the usual way to make a general model larger. He said companies need agents built around their own operational requirements, that respect boundaries, that show evidence a person can check, and that stay cheap enough to deploy. File the name, the title, and those sentences as the release’s. A quote is not a measured cost, and it is not a customer list.

The career lines the release prints, and where they stop. Ruppersberg began his career in a Nobel Prize-winning research team. The page does not name the prize, the year, or the lab. Do not add them. He later became a full professor and a department head. The page does not name the university or the department. He published in Nature and Science and filed numerous patents. Those journals and the patent line are the release’s. This desk did not open the papers or the filings. He later founded several companies, including Mentasys, which the release calls an AI product-search pioneer, acquired by Lycos in 2006. 2006 is the page’s year. This desk did not read the Lycos announcement.

What is still ahead. Ammonix says it will soon release AmmonixCode, a source-available coding platform for building agents with this architecture. Source-available means the code can be read, under a license the company sets. It is not the same phrase as a fully open license, and the page does not name the license. Soon is the company’s word. The page does not print a date. Do not add one. AmmonixCode will be free for research and evaluation. Commercial and enterprise licenses will be available. The page prints licensing@ammonix.ai for those licenses. The about line points readers to www.ammonix.ai. The page does not name a customer, a price, or a download that is live today.

Plain English for the rest of the card: System One = the fast specialized decision layer. System Two = the language model for reasoning and interaction. specialist agent = software aimed at one job, not a general chat model. a few hundred examples = the company’s line for how much work it says is enough to reach expert level. memory = what the agent updates so it improves. auditable = a person can check the recommendation against examples. ask rather than guess = if the agent does not know, it asks a person. local hardware = the computers on site. ECG = a heart tracing. GPT-6 = the OpenAI model named in the comparison. The scores, the cost, and the speed are not printed. Cortex = Ruppersberg’s earlier medical AI company. up to $300 million = the ceiling the release puts on the January 2025 Boston Scientific deal. AmmonixCode = the coding platform the company says is coming, free to try for research, paid for commercial use. This filing is the 22 Sep stealth exit.

PRIMARY here: Ammonix’s 22 Sep 2026 PR Newswire release, “Ammonix Launches a New AI Architecture for High-Stakes Specialist Agents That Outperform Frontier Models at a Lower Cost,” stamped 09:00 ET and datelined Blonay and Wilmington — Tier A PRIMARY, the company’s own record. The System One and System Two split, the operational-data line, the few-hundred-examples line, the memory update, the auditable examples, the ask-a-person rule, the local-hardware line, the four applications, the GPT-6 comparison, the Ruppersberg title, the Cortex sale for up to $300 million in January 2025, the quote, the Nobel-team sentence with no prize named, the professor line with no university named, the Nature and Science line, Mentasys and the 2006 Lycos sale, and the AmmonixCode plan are the release’s. NOT claimed: that this desk ran an agent, counted the examples, reran the GPT-6 comparison, saw a Cortex check, named a Nobel prize, a university, a license, a price, a customer, or a ship date for AmmonixCode, a stock tip, or investment advice. Distinct from the already-filed heidi-340m, ramona-optics-25m, and rabbit-os3.