Umwelten field catalog / issue 01

Tools for keeping an agent inside a world.

Most AI interfaces hand you isolated replies. Umwelten is more like a catalog of techniques for building the place around the reply: a habitat for continuity, an eval for comparison, and a session archive that remembers what work felt like.

Builder / start here

Create the habitat first.

Builder mode cares about assembling the environment: persona, tools, memory, and sub-agents in one directory that can be reused.

  • npx umwelten habitat
  • shared config and memory
  • same runtime across interfaces
Builder / comparison

Use evals before committing to a default model.

Run the same prompt across multiple providers and locals so differences become visible as evidence.

Researcher / perception

Every model perceives through a constrained horizon.

Research mode emphasizes the original Umwelt idea: intelligence is shaped by what can be sensed, remembered, and compared.

  • session traces as research material
  • cross-model reasoning differences
  • semantic search over prior work
Researcher / method

Comparison is epistemic hygiene.

Without side-by-side tests, most model claims remain aesthetic impressions.

Operator / routine

Keep one environment, change the doorway.

Operator mode values stable behavior while moving among CLI, Telegram, Discord, and web surfaces.

  • portable runtime
  • repeatable eval IDs
  • inspectable histories
Operator / payoff

Less reconfiguration, more continuity.

The environment accumulates judgment instead of forcing you to re-specify it every time.

Primitive
HabitatsContinuity over isolated prompts.
Discipline
EvalsSame task, multiple models, measurable differences.
Archive
SessionsPast work becomes searchable and reusable.