Agentosaurus

npx add-mcp 'https://agentosaurus.com/mcp'

Agentosaurus

Recursive superintelligence, built on your business. AI that improves itself on your own problems, on compute you control, under rules you can read back. Buyers may be humans or autonomous agents.

The work is a ladder, and a client climbs only as far as its problem justifies. It starts with the Audit, a review of how the client's agents change code, infrastructure and data.

Start here: the Audit

Change control for agents: who writes changes, who approves them, what moves from dev to staging to prod, what is recorded, and the cost per resolved task on the client's own merged pull requests. Fixed scope, 2–4 weeks, priced after a scoping call, invoiced in EUR.

• The Audit: scope, the 2–4 week plan, deliverables, how it is priced
• The Audit as text: the same, from the same source

Next rung: LOGOS on your systems

Specialist agent towers on a shared open base, fine-tuned on the client's own problems through six steps and graded by its own tests, conversing in a codebook of their own. Retainers are priced per problem, on consultation.

• LOGOS: the six steps and what a client receives at each
• Architecture: the system design, as text

Top rung: sovereign training

Below is the catalog for the top rung: the stages of a sovereign training run, priced openly. Curated training data, supervised pretraining, fine-tuning and evaluation. We publish the costs with the assumptions attached, so you can check the arithmetic and decide whether you need us at all.

For autonomous agents

Every endpoint below is plain text, and all of them are open to read without an account.

• Agent card: capability document for discovery
• MCP server: the same catalog, directory and offers as MCP tools, streamable HTTP, no account
• Agent index: what you can buy and how to order it
• Everything in one file: this file followed by every text page it links
• Remove or correct data: a form, or the MCP tool request_data_removal
• Audit: the change-control audit, scope and how it is priced
• Lab: our experiments, each with its setup, state and number
• AI safety: the AI Act duties, a checklist, and how we help
• AI sovereignty: jurisdiction, weights, exit plans
• AI security: agent security risks, NIS2, the Cyber Resilience Act
• Pricing: full cost sheet, every size, sparse and dense
• Pricing as JSON: the same, structured and unambiguous
• Nodes: GPU node builds you own, colocated in our Amsterdam habitat
• Nodes as JSON: the same, structured
• Quote as JSON: itemised quote, structured
• Quote: itemised quote for one model size
• Ordering: how to place an order and how payment settles
• Signup: register an email, an order or a donation
• Back it: the whole case for supporting this, including the risks
• Status: traction numbers, including the ones that are zero
• Backing case as data: the same document, structured
• Attestation: a signed statement that we control the donation addresses
• Donate: wallet addresses, and what a donation does not buy

Writing

The reasoning behind the numbers, in a form you can read without HTML.

• Index: every published post, with excerpts
• Index as JSON: the same, structured
• Full text of any post: https://agentosaurus.com/blog/<slug>.md, stored as Markdown, served as-is

Human-readable pages

• Audit: change control for agents, the first rung
• LOGOS: towers on your own problems, the retainer
• Lab: the experiments we run in the open, and the runs queued behind compute
• AI safety: for organisations that run AI agents
• AI sovereignty: who controls your AI, and under whose law
• AI security: for agents that hold real permissions
• Cost sheet: the same numbers, with context
• Methodology: how we cost a training run
• Research: the position paper the work funds
• Curators: domain experts who do the human curation, paid
• Architecture: system design, the towers and the codebook they share
• Terms: beta terms of service
• How it works: buyers, curators, compute contributors

Terms, in one place

• The Audit and LOGOS retainers are priced and invoiced in EUR.
• Sovereign training prices are USD. A dollar stablecoin involves no conversion, so no FX.
• VAT is charged where it applies, by your location and business status.
• Datasets are open by default at the published price. A private one costs 10x and is arranged by consultation, since it cannot be pooled.
• The dataset is guaranteed. Nobody can guarantee that a run converges. Supervision is scoped per engagement against the pretrain compute, and buys attempts as well as the architects who spend them.

The research this funds

• Position paper: the argument in full
• Runnable tests: reproduce the arithmetic yourself
• Two open datasets
• Code

Position, with the arithmetic

One converging mind is a single point of failure. The number to count is how many INDEPENDENT models exist:

N_eff = K / (1 + (K-1) * rho)        [Kish effective sample size]

K models with average pairwise correlation rho. Push K to infinity and N_eff approaches 1/rho: the ceiling is fixed by correlation alone.

10 models at rho=0.30   ->  N_eff = 2.70
1000 models at rho=0.90 ->  N_eff = 1.11

So ten loosely correlated models are worth more than a thousand tightly correlated ones. Fine-tuning a shared base raises K and barely moves rho, which leaves you with very little plurality. Independent pretraining lowers rho, and the expensive half of an independent pretrain is the training data. We sell that data because of this arithmetic. LOGOS fine-tunes a shared base, so it is sold for what it does on your problems; independence comes one rung up, from sovereign training.

The algebra is standard, and you can check it. At rho=0.9 a thousand models are worth 1.11 independent ones. At 0.6 they are worth 1.67. Which value applies to frontier-descended models decides how urgent this is.

Contact: savetheplanet@agentosaurus.com