Umwelt Engineering: Designing the Cognitive Worlds of Linguistic Agents

Rodney Jehu-Appiah

For a language model, the available language is not a transparent medium through which cognition passes — it is the cognition. Change the available language and you change the cognition itself.

Read the paper (PDF)

What is Umwelt Engineering?

Jakob von Uexküll coined Umwelt to describe the perceptual world an organism's biology makes available to it. A tick's world contains butyric acid, temperature, and tactile density — nothing else exists for it.

A language model's cognition unfolds entirely in the token stream. The words don't describe the thinking; they are the thinking. Umwelt engineering is the deliberate design of this linguistic cognitive environment — a third layer in the agent design stack, above prompt engineering (what the agent is asked) and context engineering (what the agent knows).

Key Findings

Experiment 1 — 4,470 trials across three models (Claude Haiku 4.5, GPT-4o-mini, Gemini 2.5 Flash Lite), three conditions, and seven reasoning tasks:

• Removing "to be" (E-Prime) improves ethical reasoning by +15.5pp and causal reasoning by +14.1pp, while degrading syllogisms by -3.4pp
• Removing possessive "to have" (No-Have) improves ethical reasoning by +19.1pp and classification by +6.5pp, with near-perfect compliance (<1% violation rate vs. E-Prime's 52%)
• Effects are model-dependent: Gemini benefits enormously from E-Prime (+42pp on ethical reasoning), GPT-4o-mini collapses on epistemic calibration (-27.5pp), Haiku shows small effects. Cross-model correlations are negative — evidence that different models occupy different native Umwelten.

Experiment 2 — 16 linguistically constrained agents on 17 debugging problems:

• No constrained agent outperforms the control individually
• A 3-agent ensemble selected for linguistic diversity achieves 100% ground-truth coverage vs. 88.2% for the control
• The constraints make agents different from each other in useful ways, even when they don't make any single agent better

Repository Structure

paper/                  # LaTeX source, bibliography, compiled PDF
tasks/                  # 130 task items across 7 reasoning categories
prompts/                # System prompts for control, E-Prime, No-Have conditions
scripts/                # Experiment runner, scoring, statistical analysis
results/                # Raw trial data (JSONL) and scored exports (CSV)
analysis/               # Statistical reports and visualizations
config/                 # Experiment configuration

Reproducing the Experiments

pip install anthropic openai google-genai pyyaml

# Set API keys as environment variables
# ANTHROPIC_API_KEY, OPENAI_API_KEY, GEMINI_API_KEY

# Run the full experiment (~5.5 hours, ~$15 in API costs)
python scripts/run_experiment.py

# Score results
python scripts/score_results.py

# Statistical analysis
python scripts/analyze_statistics.py

Citation

Paper is forthcoming on ArXiv. In the meantime:

Jehu-Appiah, R. (2026). Umwelt Engineering: Designing the Cognitive Worlds
of Linguistic Agents. https://github.com/rodspeed/umwelt-engineering

License

Code: MIT. Paper: All rights reserved.