An agent improves itself: one hour on one graphics card
Our coding agent spent an hour on its own source code, with Qwen3.8 on an RTX 5090. 13 of 16 iterations went into the code.
At the top are articles that grew out of our own work: building the coding agent, the Sentinel validation layer and projects with a regulatory angle. Below follows a fundamentals section for reference. Every article appears in German and English.
Our coding agent spent an hour on its own source code, with Qwen3.8 on an RTX 5090. 13 of 16 iterations went into the code.
After 24 days our coding agent has fully designed a humanoid robot. It walks in simulation without a fall and grasps. With the measurements.
Our coding agent manages its context window in two tiers. The run logs showed placeholders that invited repetition, and one overflow.
Three Sentinel rules are valid KQL and stay silent in operation. Our check reports all three as errors. With the findings quoted verbatim.
Our humanoid robot exists as a design and a simulation. The simulation follows three rules; for one of them we withdrew our best results.
Qwen3.8-27B on an RTX 5090 with speculative decoding: 23.7 becomes 84 to 115 tokens per second. Graphics memory limits depth and context.
A coding agent reported done without ever running its code. Since then the last run decides, and plan steps need a green gate command.
Qwen3.8 runs on a single graphics card and trails the frontier by months. Regulation targets providers, and open models barely have one.
A prompt alone does not secure a coding agent with shell access. We secure our own agent with a permission engine, a verified sandbox and undo.
I counted the programme of World Summit AI 2026 in Amsterdam: providers, researchers and supervisors spoke about risk and AI regulation.
CardinalOps found that 13% of rules in production SIEMs never fire. Sentinel rules can be checked deterministically against Microsoft’s KQL parser.
NIS2 requires an early warning within 24 hours of awareness. The hurdle is detecting the incident. What Article 23 requires and what you build for it.
An LLM writes a detection query in seconds. Whether it is correct only shows in a generate-validate-refine loop that runs it against the schema and sample data.
Million-token context windows did not replace RAG. Long context stays affordable only when the prompt prefix is identical for every query.
Annex III of the EU AI Act does not make your AI high-risk by itself. The Article 6(3) filter can rule it out, and profiling blocks the filter.
An AI system returns a different number on every measurement. Acceptance becomes testable when the contract owes a threshold on a frozen data set.
Article 12 of the EU AI Act names no log fields for an LLM. Which corpus produced a RAG answer is proven only by the index snapshot. With JSON schema.
In most vector databases delete() only writes a tombstone. What Art. 17 GDPR requires and how you prove the erasure.
Almost any AI tool your staff use triggers co-determination under § 87 BetrVG. The four provisions that apply and what a works council agreement covers.
How to build AI on personal data in line with the GDPR: which duties apply, where projects fail, and which architecture anchors the duties in the system.
Hardware for local LLMs: choose a model size, estimate GPU memory (VRAM), save with quantization, and know when on-premise beats the cloud.
Introductions to the terms and methods the articles above build on.