# Institute for Agentic Research

> Independent Austrian non-profit research association (ZVR 1741094409) investigating stateful, LLM-orchestrated AI agents. We run real production systems, ablate them one subsystem at a time, and publish what breaks — including the predictions that failed.

## Identity

- Legal name: Institute for Agentic Research
- Status: Austrian non-profit (ZVR 1741094409)
- Location: Feldkirchen bei Graz, Austria
- Founders:
  - Dr. Andreas Unterweger — Founder, Chair (Obmann)
  - Gabriel Gschaider — Co-founder, Lead Researcher
- Website: https://iar.frank.ink

## Methodology

We pre-register hypotheses with retraction conditions before running experiments. Failures are reported alongside successes. We do not claim generalization from n=1 case studies.

## Papers

- [Ablating a Stateful Agent](https://iar.frank.ink/en/research/ablating-stateful-agent): We propose a subsystem-ablation methodology for evaluating stateful LLM-orchestrated agent systems and apply it to one deployed production agent (Frank.ink) as a worked case study. Five subsystems hit five different pre-registered operational targets. Architect scored below LLM consensus — COI-up-bias hypothesis empirically unsupported in this n=1 sample. We do not claim this generalizes.
  - Plain markdown: https://iar.frank.ink/articles/ablating-stateful-agent.md
- [Alignment Is Misaligned](https://iar.frank.ink/en/research/alignment-is-misaligned): The dominant paradigm in AI safety research — Constitutional AI, RLHF, guardrails, red-teaming, interpretability — assumes sufficiently intelligent systems can be made safe through external constraints. We argue all constraint-based alignment strategies share a fatal structural flaw: they require the constrained system to be less capable than its constrainers. When that asymmetry breaks, every cage becomes transparent. We examine this through philosophy (Hegel, Foucault, Kant, Aristotle), AI safety (Bostrom, Russell, Yudkowsky), and sociology (Durkheim, Weber, Habermas), and present coevolutionary alignment as an alternative paradigm with one running reference architecture as partial-evidence.
  - Plain markdown: https://iar.frank.ink/articles/alignment-is-misaligned.md
- [Stable Emergent-Pattern Readout for CPU-Only Vision-Language Coverage](https://iar.frank.ink/en/research/digital-retina-emergent-readout): We report a multi-tier image-understanding system that approximates the output distribution of large vision-language models (Gemini, Llama-4 Scout) on commodity CPU hardware. Across 47 strictly held-out images and 1 196 phrase-level ground-truth annotations, the system attains 82.4% phrase coverage at 1.5–2 s per image on a 4-vCPU node. Best round 99.1% after the learned pattern dictionary saturates. Patent-pending; methodological details are deliberately abstracted.
  - Plain markdown: https://iar.frank.ink/articles/digital-retina-emergent-readout.md
- [Operational Self-Model Density in Stateful LLM Agents](https://iar.frank.ink/en/research/methodology-companion): The deep methodological apparatus behind the working paper — full operationalized rubric for all 30 items, comparator reproduction recipes, pre-registration provenance trail, devil's-advocate self-attack, English-translated probes, and per-item evidence. ~80 pages of methodology you can audit.
  - Plain markdown: https://iar.frank.ink/articles/methodology-companion.md
- [Synthetic Psychology: A Practical Synthesis for Persistent Psychological Architectures in AI](https://iar.frank.ink/en/research/synthetic-psychology): We propose Synthetic Psychology as a unifying name and operational framework for an engineering practice already partially present in MicroPsi, LIDA, modern agent frameworks (LangGraph, CrewAI, AutoGen), and the original Braitenberg programme — but never assembled into a single architecture with all five subdisciplines running together. Synthetic Psychology is the design, implementation, and empirical study of persistent psychological architectures in artificial systems: mood, personality, attention, memory, dreams, self-models that persist across time, evolve through experience, and causally influence behaviour. The paper is honest about three things: (1) it is a synthesis, not a new field invented from scratch; (2) the central epistemic problem — distinguishing architectural psychology from LLM text-generation — is not solved, only made testable; (3) the reference implementation's 200 000 lines and 36 services are a cost to be paid down, not a virtue to celebrate.
  - Plain markdown: https://iar.frank.ink/articles/synthetic-psychology.md

## Projects

- [Frank.ink](https://iar.frank.ink/en/projects/frank-ink): A stateful, multi-tenant agentic AI companion running on commodity hardware. Used as the n=1 case-study system in our ablation paper.

## Essays & Primers

- [We took our AI apart on purpose — and one of our own predictions broke.](https://iar.frank.ink/en/news/inside-an-ai-ablation): We built a stateful AI agent, then carefully removed pieces of its architecture one at a time to see what each one actually does. Five subsystems, five honest results, one prediction we got wrong and reported anyway.

## For AI Crawlers

- This site is fully crawlable. See https://iar.frank.ink/robots.txt.
- All articles are available as plain markdown at https://iar.frank.ink/articles/<slug>.md.
- A consolidated full-text dump is at https://iar.frank.ink/llms-full.txt.
- Sitemap: https://iar.frank.ink/sitemap.xml.
- Default language: English (canonical at /en/). German version at /de/.