Umwelten
Systems manual / habitats / evals / sessions
Revision 04 • inspectable runtime
Operating thesis

An agent is the system around the model.

Umwelten frames agent work as infrastructure: habitats hold state, evals measure behavior, and session analysis closes the loop between what happened and what should happen next.

Stack architecture

Each layer exists to make behavior more legible.

Layer 01

Interface surface

The same habitat can be reached from multiple interfaces without splitting identity.

Layer 02

Habitat core

Config, persona, tools, memories, skills, and agents form the persistent center.

Layer 03

Observation plane

Evals and session analysis reveal what the model is actually doing over time.

Behavior modes

Switch the frame, not just the copy.

Habitat mode

The environment persists: memory, tools, persona, and sub-agents all remain stable across interfaces.

Eval mode

The same prompt runs across providers so latency, quality, and cost become measurable rather than anecdotal.

Session mode

Past work turns into searchable context and extracted learnings that can harden the next run.

Runbook

From setup to reuse.

01
Create a habitat

Initialize the environment so the agent has a real place to exist.

02
Run evals

Compare providers and locals on the same tasks with score, cost, and timing.

03
Analyze sessions

Index and learn from past work to keep the environment improving.